<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Raell Dottin on Medium]]></title>
        <description><![CDATA[Stories by Raell Dottin on Medium]]></description>
        <link>https://medium.com/@raell.dottin?source=rss-fd9ac23f0a87------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Ig2pglH_LlfYbDuMJgCvFQ.jpeg</url>
            <title>Stories by Raell Dottin on Medium</title>
            <link>https://medium.com/@raell.dottin?source=rss-fd9ac23f0a87------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 09:22:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@raell.dottin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[I Met a Vibe Coder Who Built His Own Trading Platform]]></title>
            <link>https://medium.com/@raell.dottin/i-met-a-vibe-coder-who-built-his-own-trading-platform-0e4f6c4befc9?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e4f6c4befc9</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-11-16T17:01:02.922Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>At a Tailscale and Speakeasy event, I met someone who knew little about programming but had built a trading platform that apparently changed his finances.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oKA33VyEA6R9Vmnr" /><figcaption>Photo by <a href="https://unsplash.com/@productschool?utm_source=medium&amp;utm_medium=referral">Product School</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>On October 14, I attended The Allowlist, a Tailscale and Speakeasy event at Tanner Smith’s in New York City. The gathering brought together people interested in artificial intelligence, software development, and how AI systems should connect to other applications. There were drinks, conversations, and no presentations to sit through, which suited me just fine. I could spend the evening talking to people whose projects I might otherwise encounter only through GitHub or a Discord server. [1][3]</p><p>The organizers had their own technical news. The previous day, Speakeasy and Tailscale had announced a partnership that allows companies to connect AI agents to internal tools over private networks. It was relevant to the event’s focus on privacy, although that announcement wasn’t what stayed with me afterward. [4]</p><p>I met several interesting people, but one person held my attention longer than anyone else. He had built his own trading platform, and he described financial success that sounded far beyond what I had imagined someone could accomplish with a personal software project. What surprised me even more was how little conventional programming he appeared to know.</p><p>I’ve been building software with AI tools myself, so I finally had an opportunity to compare experiences with someone approaching development from a very different background. One of the projects he described reminded me of something I was building. That made the conversation considerably more interesting than another debate about which AI model performed best on a benchmark.</p><p>I wanted to understand how he worked, which models he trusted, and how he decided that software written largely by AI was ready to use. His trading platform was impressive as he described it, but the process that allowed him to build it was what I wanted to investigate.</p><h3>Building software without knowing much programming</h3><p>The term <em>vibe coding</em> became popular in 2025 to describe building software by telling an AI model what you want and letting it generate the implementation. It is frequently associated with accepting generated code without carefully examining every line. In practice, people now use the phrase for activities ranging from simple experiments to maintaining applications that have become considerably more complicated. [5]</p><p>I’ve never considered programming knowledge an absolute requirement for having a useful software idea. People who understand a business process often know exactly what they want from an application, even if they cannot explain how to implement it. AI coding tools have made it substantially easier for those people to attempt the implementation themselves.</p><p>A trading platform is an ambitious choice of project, though, because its responsibilities can extend well beyond displaying information on a screen. Depending on the application, it might need to retrieve market data, calculate indicators, maintain records of positions, communicate with a brokerage, or submit actual orders. Each additional capability introduces ways for an incorrect assumption to cause problems.</p><p>I didn’t inspect his source code, use the platform, or establish exactly how much of the trading process it automated. I also didn’t examine his account statements. The improvement in his standard of living was something he described, not a financial result I independently verified.</p><p>I mention those limits because I was interested in learning from him without pretending I had performed a technical audit. I had met someone who appeared to be accomplishing a great deal with AI-assisted development, and his account was interesting enough to warrant more questions.</p><p>I started probing him the way I might use Nmap, a network-scanning tool, to discover which services are available on a machine. I wanted to know how the pieces connected and what was happening behind the interface. [17]</p><p>His first answer gave me a useful place to start: he said he spawned a subagent for every prompt.</p><h3>Every prompt gets another worker</h3><p>A subagent is a separate AI worker assigned a task by another AI agent. Instead of having one conversation investigate a problem, modify the code, and remember everything it encountered, a coordinating agent can delegate some of that work to another worker. The subagent operates with its own context, meaning the information available to it while completing the assignment, then returns its findings or completed work.</p><p>Anthropic documents this approach in Claude Code as a way to prevent side investigations from overwhelming the main conversation. One worker might examine how authentication works in an application while another investigates the tests. The coordinating agent can then use those results without retaining every intermediate file and error message encountered during the investigations. [6]</p><p>I could appreciate why someone would use that approach. Once a software project grows beyond a handful of files, an AI coding session can accumulate an extraordinary amount of information. The agent reads source code, examines errors, experiments with changes, and revisits earlier decisions. Eventually, managing that accumulated context can become part of the work itself.</p><p>Delegating independent investigations can help, particularly when the results are straightforward to summarize and the workers do not need to edit the same files simultaneously.</p><p>What intrigued me was how frequently he said he did it. He wasn’t describing occasional delegation for difficult assignments. As I understood him, creating subagents was part of his normal response to prompting.</p><p>There are costs to that approach, and they aren’t limited to the amount paid for additional AI model usage. Every worker needs an assignment, and its results must be integrated into the wider project. Parallel workers can produce conflicting changes or rely on assumptions that another worker has not verified. Anthropic’s own guidance cautions against excessive delegation when a direct operation would be simpler. [7]</p><p>I don’t know whether his approach was faster or more reliable than mine. We hadn’t run the same assignment against the same project under controlled conditions, and I didn’t have enough information about his operating costs or failure rates to make that comparison.</p><p>Nevertheless, the idea gave me something worth examining in my own workflow. I tend to organize coding agents around assignments and responsibilities, with delegation serving a particular purpose. His account suggested a more aggressive approach, where isolating each assignment could be valuable enough to justify additional workers on a regular basis.</p><p>The benefit depends partly on what happens after those workers return. Producing changes through more agents may increase the amount of work a project can attempt, but those changes still need to be reconciled and evaluated. When he mentioned that he used automated code reviews, I began to understand how he might have organized that part of the process.</p><h3>Who reviews code when the author can’t explain all of it?</h3><p>A code review is an examination of proposed software changes intended to identify bugs, unsafe behavior, or other problems before the changes become part of an application. Developers often review a <em>pull request</em>, a set of proposed changes submitted for examination before being merged into the project’s main codebase.</p><p>I knew the value of code review, but our conversation led me to investigate how different AI coding tools provided it.</p><p>Codex has a /review command that examines changes in a project and reports potential problems without modifying the files. It can review work that hasn&#39;t been committed or compare a development branch with another branch. OpenAI also provides code-review workflows for pull requests. [8]</p><p>I subsequently discovered that Claude Code had comparable functionality. Anthropic provides a /code-review plugin that uses multiple specialized agents to examine proposed changes. Its separate organizational Code Review service became generally available on October 15, the day after the event, for eligible Team and Enterprise customers. [9][2]</p><p>I’d been looking for an equivalent without realizing how it was distributed. That discovery was useful on its own, but it also made me think about the assumptions behind reviewing AI-generated software.</p><p>A conventional review often assumes the developer understands a change well enough to explain why it was made and how it is supposed to behave. That assumption becomes less dependable when the person directing the work has limited programming experience and relies on generated explanations to understand the implementation.</p><p>Independent AI reviewers can help identify incorrect conditions, missing error handling, or code that does not satisfy the requirements. Their conclusions still need examination, however. Multiple agents can make the same mistaken assumption, and tests can pass while leaving an important scenario untested.</p><p>Anthropic’s engineering account of its own code-review practices describes combining automated reviews with tests, operational monitoring, and the ability to reverse changes when problems appear. The account is useful because the company treats automated review as part of an engineering process that continues after the code has been accepted. [10]</p><p>Consider a hypothetical trading application that submits an order to a brokerage. If the connection times out immediately afterward, the application may not know whether the order succeeded. Retrying without checking the order’s status could create a duplicate transaction. A test that merely checks whether the application can submit a valid order would not establish that it behaves safely in this situation.</p><p>I don’t know whether his platform has that vulnerability or whether he already tests for it. The example illustrates why I would want to know more about the checks surrounding a financially consequential application, regardless of whether the implementation came from a human programmer or an AI model.</p><p>His use of reviews was encouraging because it suggested he was doing more than prompting an agent and accepting whatever appeared. I still wanted to know how he arrived at the trading decisions the software was meant to support.</p><h3>Learning to trade through Jane Street accounts</h3><p>He told me that his introduction to trading involved following accounts associated with Jane Street employees on X, formerly Twitter.</p><p>Jane Street is a quantitative trading firm, meaning it uses mathematics, research, and software extensively in its trading operations. The firm also makes technical and educational material available publicly. [11]</p><p>I found his origin story interesting because it described a route into a specialized field that didn’t begin with a conventional programming education. Someone could study public explanations of financial markets, develop an understanding of the decisions traders make, and use AI tools to help translate those ideas into software.</p><p>I didn’t record the names of the people he followed or the particular posts that influenced him. I also have no evidence that anyone disclosed confidential trading strategies. Learning from public commentary by people who work at a quantitative trading firm is different from acquiring the firm’s proprietary research.</p><p>Even with those uncertainties, his account was plausible. It resembled the way many people teach themselves technical subjects: follow experienced practitioners, read explanations, experiment, and gradually acquire the vocabulary needed to work through more difficult problems.</p><p>I’ve done versions of that throughout my own career. What has changed is how quickly someone can attempt to implement an idea after developing a basic understanding of it. AI can provide much of the initial implementation, leaving the person directing the work to decide whether the result reflects what they intended.</p><p>We also compared AI models during the conversation. I didn’t preserve a reliable list of the models he named, so I won’t reconstruct one from memory. What stayed with me was that he had made deliberate choices about his tools, despite programming not being his original field of expertise.</p><p>Eventually, we got around to how he actually approached trading. That was where the conversation moved from building software to evaluating the decisions the software was supposed to support.</p><h3>Thirty minutes, MACD, and SMA9</h3><p>He described avoiding trading during the first 30 minutes after the market opens and using two technical indicators: MACD and SMA9.</p><p>For stocks traded during the regular New York Stock Exchange session, the market opens at 9:30 a.m. Eastern, so waiting half an hour would ordinarily mean beginning no earlier than 10:00 a.m. The opening period can experience substantial trading activity and price volatility as orders and reactions to overnight information reach the market. New York Stock Exchange research has examined the unusual trading conditions during that period. [12]</p><p>I didn’t establish exactly which assets his platform traded, so the New York Stock Exchange schedule is an illustration of how the rule applies to U.S. equities. Other markets operate on different schedules.</p><p>Waiting until the initial rush has passed is an understandable personal rule. Whether it improves a particular strategy depends on the assets being traded, the way positions are entered and exited, and what opportunities the rule excludes.</p><p>The technical indicators require some explanation. MACD stands for <em>Moving Average Convergence/Divergence</em>. It compares two moving averages of historical prices to help identify changes in market momentum. A moving average smooths price fluctuations so a trader can observe a broader pattern. The conventional MACD calculation uses averages covering 12 and 26 periods, together with a nine-period signal line that helps traders interpret changes in the indicator. [13]</p><p>SMA9 means <em>nine-period simple moving average</em>. It is the arithmetic average of nine recent price observations. On a one-minute chart, that represents nine minutes of observations; on a daily chart, it represents nine trading days. The interval changes how quickly the indicator responds to price movements and what information it provides. [14]</p><p>Neither indicator predicts future prices reliably on its own. Both summarize previous price behavior, which may help a trader decide whether a particular set of conditions exists.</p><p>I was especially interested in how he described these practices as risk management. Avoiding the opening 30 minutes limits exposure to a particular trading period. MACD and SMA9 may inform decisions about when to enter or leave a position. They do not determine how much money a strategy could lose on a trade, how it handles a prolonged losing streak, or whether a technical failure could create an unintended position.</p><p>Those questions require additional controls, and they become more important when software can communicate directly with a brokerage. A strategy might calculate its indicators correctly and still lose money because its positions are too large. It might appear profitable when tested against historical prices but perform poorly with real orders because of transaction costs, changing market conditions, or differences between expected and actual execution prices.</p><p>The Securities and Exchange Commission warns that day trading can result in severe losses, including losses amplified by borrowed money. Investor guidance from the SEC and other regulators also warns against relying solely on AI-generated information when making investment decisions. [15][16]</p><p>None of that establishes that his approach was unsuccessful. He may have had additional controls that never came up during our conversation, and he may have spent considerable time testing the strategy. I didn’t establish enough about the platform to evaluate either possibility.</p><p>What I could say was that the indicators he mentioned explained part of how he interpreted market behavior. They didn’t independently explain his reported financial success or establish whether the strategy would continue performing well under different market conditions.</p><p>That observation influenced how I thought about the engineering side of his project. With coding agents, I usually want evidence that an implementation satisfies its requirements and survives relevant tests. A trading application adds another evaluation problem: the software may perform exactly as designed while the trading rules themselves produce financial losses.</p><p>This is why I kept returning to the relationship between his development process and his trading process. Automated code reviews could help him find programming errors, but evaluating a trading strategy requires different evidence, including the treatment of losses, execution costs, and performance across changing conditions. I would have liked to know how he connected those two kinds of evaluation.</p><h3>What I took home</h3><p>By this point, I had collected quite a few ideas to investigate. I wanted to reconsider when I spawn subagents, how I organize independent code reviews, and how much I rely on generated explanations when evaluating generated software. I was also curious about the models he preferred and the way he connected his market research to his development workflow.</p><p>I wasn’t looking to abandon the way I build applications. Some of my habits come from years of working with systems that have to remain maintainable after the original developer has moved on. Those experiences have made me cautious about changes that are difficult to explain, reproduce, or reverse.</p><p>His experience suggested another route into software development was becoming practical. Someone could start with a problem they understood, use AI to construct an initial application, and gradually acquire technical knowledge while working through real requirements.</p><p>There is a strong argument for that approach. Requiring everyone to learn programming in the conventional order before attempting a useful project excludes people whose expertise lies elsewhere. An experienced trader might understand the desired behavior of a trading tool better than an experienced programmer who knows very little about markets.</p><p>It also leaves a responsibility that cannot be delegated away simply by creating another agent. Someone has to decide which results are acceptable and what evidence is sufficient before relying on the system. For ordinary software, that might mean checking how it handles errors, stores information, or recovers from failures. For a trading platform, the consequences extend into actual financial decisions.</p><p>I would have liked to continue questioning him. There was more I wanted to learn about his platform, his review process, and how he decided that generated software was ready to use. Unfortunately, I eventually realized that I’d had three glasses of Greek white wine, and heading home seemed considerably more sensible than conducting another round of technical interrogation.</p><p>I left with several things to investigate and an unexpected appreciation for the conversation. I had gone to a technology event hoping to meet interesting people, and I came away with useful questions about my own development process.</p><p>I still don’t know exactly how successful his trading platform is, or how much of that success can be attributed to the software itself. I’d enjoy comparing notes again, preferably before the third glass of wine.</p><h3>References</h3><p>[1] Speakeasy. “Registration approved for The Allowlist” and “New message in The Allowlist.” Event correspondence received October 14, 2026. Private emails retained by the author.</p><p>[2] Anthropic. “Code Review Is Now Generally Available.” October 15, 2026. <a href="https://claude.com/resources/articles/code-review-is-now-generally-available">https://claude.com/resources/articles/code-review-is-now-generally-available</a></p><p>[3] Speakeasy and Tailscale. “The Allowlist.” Public event listing, October 2026. <a href="https://luma.com/po85yuqq">https://luma.com/po85yuqq</a></p><p>[4] Speakeasy. “Speakeasy and Tailscale Bring Private Networking to Enterprise MCP.” October 13, 2026. <a href="https://www.speakeasy.com/blog/partnering-with-tailscale">https://www.speakeasy.com/blog/partnering-with-tailscale</a></p><p>[5] Cloudflare. “What Is Vibe Coding?” Cloudflare Learning Center. <a href="https://www.cloudflare.com/learning/ai/ai-vibe-coding/">https://www.cloudflare.com/learning/ai/ai-vibe-coding/</a></p><p>[6] Anthropic. “How and When to Use Subagents in Claude Code.” April 7, 2026. <a href="https://claude.com/resources/articles/subagents-in-claude-code">https://claude.com/resources/articles/subagents-in-claude-code</a></p><p>[7] Anthropic. “Prompting Best Practices.” Claude Platform documentation, section on subagent orchestration. <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices">https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices</a></p><p>[8] OpenAI. “Code Review.” ChatGPT Learn documentation. <a href="https://learn.chatgpt.com/docs/code-review">https://learn.chatgpt.com/docs/code-review</a></p><p>[9] Anthropic. “Code Review.” Claude Code plugin documentation. <a href="https://claude.com/plugins/code-review">https://claude.com/plugins/code-review</a></p><p>[10] Addy Osmani, Morgan Westlee Lunt, and Qing Wang. “How the Claude Code Team Does Code Review.” October 15, 2026. <a href="https://claude.dev/blog/solving-code-review-at-anthropic/">https://claude.dev/blog/solving-code-review-at-anthropic/</a></p><p>[11] Jane Street. Official company website, including information about the firm’s trading activities, research, and technical publications. <a href="https://www.janestreet.com/">https://www.janestreet.com/</a></p><p>[12] Choey Li. “Opening Auction Filling Behaviors: Opportunities Early in the Day.” New York Stock Exchange, December 2, 2021. <a href="https://www.nyse.com/data-insights/opening-auction-filling-behaviors-opportunities-early-in-the-day">https://www.nyse.com/data-insights/opening-auction-filling-behaviors-opportunities-early-in-the-day</a></p><p>[13] StockCharts ChartSchool. “MACD (Moving Average Convergence/Divergence) Oscillator.” <a href="https://chartschool.stockcharts.com/table-of-contents/technical-indicators-and-overlays/technical-indicators/macd-moving-average-convergence-divergence-oscillator">https://chartschool.stockcharts.com/table-of-contents/technical-indicators-and-overlays/technical-indicators/macd-moving-average-convergence-divergence-oscillator</a></p><p>[14] TradingView. “Simple Moving Average.” Technical indicator documentation. <a href="https://www.tradingview.com/support/solutions/43000696841-simple-moving-average/">https://www.tradingview.com/support/solutions/43000696841-simple-moving-average/</a></p><p>[15] U.S. Securities and Exchange Commission. “Day Trading: Your Dollars at Risk.” <a href="https://www.sec.gov/files/investor/pubs/daytips.htm">https://www.sec.gov/files/investor/pubs/daytips.htm</a></p><p>[16] Securities and Exchange Commission, Financial Industry Regulatory Authority, and North American Securities Administrators Association. “Artificial Intelligence (AI) and Investment Fraud: Investor Alert.” January 25, 2024. <a href="https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-alerts/artificial-intelligence-fraud">https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-alerts/artificial-intelligence-fraud</a></p><p>[17] Gordon Lyon. <em>Nmap Network Scanning: The Official Nmap Project Guide to Network Discovery and Security Scanning.</em> Nmap Project. <a href="https://nmap.org/book/toc.html">https://nmap.org/book/toc.html</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e4f6c4befc9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Testing Reflection’s Beam Against GLM-5.3 While the Beta Clock Is Running]]></title>
            <link>https://medium.com/@raell.dottin/testing-reflections-beam-against-glm-5-3-while-the-beta-clock-is-running-1eb2a59eb647?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/1eb2a59eb647</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 17:01:01 GMT</pubDate>
            <atom:updated>2026-11-15T17:01:01.958Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Former DeepMind researchers have built an open-weight AI model. I’m testing whether its efficiency, coding capabilities, and commercial terms make it worth adopting.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DDX-tiUC3NgYIOZV" /><figcaption>Photo by <a href="https://unsplash.com/@theshubhamdhage?utm_source=medium&amp;utm_medium=referral">Shubham Dhage</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I recently obtained an application programming interface (API) key for Reflection AI’s Beam model. An API allows software to communicate with another service, so I can connect Beam to development tools rather than interact with it exclusively through a website.</p><p>Reflection’s Beta plan immediately caught my attention. My account shows an allowance of 10 million uncached input tokens per day, one million output tokens, and up to 20,000 requests. The plan is free, includes API access, and resets its daily allowances at midnight Coordinated Universal Time (UTC). It is also labeled as a limited-time offering. [2]</p><p>Tokens are the units of text a language model processes. An input token represents part of the information sent to the model, while an output token represents part of its response. Uncached input refers to information that must be processed without the benefit of previously stored computation. For someone experimenting with coding agents, Reflection’s allowance provides considerable room to work.</p><p>I was pleased to have access. I was considerably less enthusiastic about the possibility of discovering an unexpected bill after the Beta program ends.</p><p>I’ve become increasingly interested in the economics of using artificial intelligence for software development. An inexpensive model that requires several attempts to complete a task can consume more resources than an expensive model that succeeds immediately. When automated agents can make hundreds of requests without direct supervision, billing arrangements become part of the system’s operational design.</p><p>Reflection presented another reason to investigate. The company was founded by former researchers from Google’s DeepMind organization, which has been responsible for some of the most consequential developments in modern artificial intelligence.</p><p>I’ve been critical of Google’s difficulty turning its research advantages into the kind of consumer products that competitors managed to popularize. Finding another company founded by people from that research ecosystem, releasing a model I could actually test, made the history more interesting.</p><p>The question I’m investigating is whether Beam can offer enough capability and efficiency to compete with GLM-5.3, a model I’ve been evaluating for development work.</p><p>There is a broader question underneath that comparison. Publicly available model weights give developers more choices about where and how artificial intelligence runs, but those choices become useful only when the model performs reliably, the cost of completing work is acceptable, and the developer understands the conditions governing the service.</p><p>Reflection’s origins explain why I wanted to try Beam. Its performance, economics, and terms will determine whether I continue using it.</p><h3>Google’s research helped create its competitors</h3><p>In 2017, researchers working primarily within Google Brain and Google Research published <em>Attention Is All You Need</em>, the paper that introduced the Transformer architecture. [5]</p><p>The Transformer uses an attention mechanism to determine how different parts of an input relate to one another. Earlier approaches to language processing often depended heavily on processing sequences incrementally. The Transformer made it easier to train large systems through parallel computation while learning relationships across sequences of text.</p><p>That architecture became foundational to modern large language models (LLMs), including systems capable of producing text, writing software, analyzing documents, and interacting with development tools.</p><p>Google had helped establish the technical direction of an industry in which other companies would later become major competitors.</p><p>For years, I’ve wondered why Google, with its research talent, computing infrastructure, and established products, did not translate those advantages into a more decisive early position in conversational AI.</p><p>The history is more complicated than suggesting Google cared exclusively about scientific research. The company introduced Bard in February 2023, before opening it to the public the following month. In April, it combined Google Brain and DeepMind into Google DeepMind, consolidating two important research organizations while pursuing more ambitious AI products. [6] [7]</p><p>Google was clearly developing consumer applications. Its scientific work and commercial development were happening simultaneously, even if competitors gained public attention faster in certain areas.</p><p>Research organizations and product teams also face different pressures. Research rewards advances that may take years to become useful. Product development introduces requirements around reliability, operating costs, accessibility, safety, and the needs of millions of users.</p><p>A company attempting to excel at both must decide how researchers, infrastructure teams, and product organizations share authority and resources. Those decisions influence what gets built, how quickly work reaches customers, and how much freedom researchers have to pursue alternative approaches.</p><p>It would be tempting to explain Reflection’s founding as a direct consequence of Google’s move toward consumer AI. I haven’t found evidence establishing that its founders left for that reason, and the chronology alone does not prove it.</p><p>What the history does establish is that expertise developed within a major research organization became part of an independent competitor.</p><p>Reflection was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both formerly at Google DeepMind. Laskin led reward modeling work for Gemini, while Antonoglou was a co-creator of AlphaGo, DeepMind’s system for playing the board game Go. In October 2025, TechCrunch reported that Reflection had raised $2 billion at an $8 billion valuation after initially concentrating on autonomous coding agents. [8]</p><p>Both founders brought experience relevant to systems that learn through feedback.</p><p>Reward modeling helps assess the quality of a model’s behavior. Reinforcement learning improves a system through feedback about its actions and outcomes. These techniques are especially relevant to coding agents, which may need to attempt a solution, examine the results, and adjust their approach.</p><p>A coding agent might modify a source file, run a test, read the failure, and try again. Its usefulness depends partly on whether it can learn from the evidence available during that sequence rather than repeatedly producing plausible-looking code.</p><p>The significance of Reflection’s founding is not that Google’s researchers necessarily understood something their former employer did not. It is that the expertise required to build sophisticated AI systems is no longer confined to the organizations in which much of that expertise developed.</p><p>Researcher mobility, outside investment, and access to computing infrastructure have helped make additional competitors possible. The resulting competition gives developers more models to evaluate, although the founders’ credentials cannot establish which model will perform best.</p><h3>What Reflection built with Beam</h3><p>Reflection introduced Beam on October 5, 2026, and released its model weights on October 16 under the Apache 2.0 license. That license generally permits use, modification, and redistribution, including commercial use, subject to its conditions. [9]</p><p>Model weights are the numerical values learned during training that influence how a neural network processes information. Releasing those weights allows developers to download and operate the model using infrastructure they control.</p><p>Beam is therefore described as <em>open-weight</em>. That does not mean every element of its development is public. Reflection has not released everything required to reproduce the complete training process, including all underlying training data.</p><p>The distinction is relevant because downloading a model and independently reproducing its creation are different capabilities.</p><p>Beam uses a mixture-of-experts architecture. According to Reflection’s model card, it contains approximately 501.3 billion parameters, with about 23.4 billion active when processing each token. It supports a context window of approximately one million tokens, allowing it to process unusually large amounts of information in a request, subject to available capacity and operational limits. [10]</p><p>A parameter is a learned numerical value within the model. A mixture-of-experts system contains multiple specialized portions of a neural network and uses a routing mechanism to activate selected portions for a particular computation.</p><p>Instead of requiring all 501 billion parameters to participate in every step, Beam activates a much smaller subset. This can reduce the arithmetic required to generate a response compared with a similarly sized model that activates every parameter.</p><p>It does not eliminate the cost of storing the complete model, transferring information through memory, processing long inputs, or operating the surrounding infrastructure.</p><p>Reflection reports that Beam was pretrained on 23.8 trillion tokens and subsequently received substantial reinforcement-learning training. The company argues that its architecture and training methods allow it to approach the capabilities of larger systems with fewer generated reasoning tokens. [11]</p><p>Reasoning tokens represent part of the computation a model uses while working through a problem before producing its final answer. For providers that charge for those tokens, reducing their consumption can lower a task’s cost, assuming the model still completes the task successfully.</p><p>Reflection’s efficiency analysis deserves a careful reading. Its published estimates use active parameter counts and generated-token totals to approximate computational work. Those estimates exclude some costs, including processing the original prompt, context-dependent attention calculations, and serving overhead. [11]</p><p>Consequently, fewer active parameters or reasoning tokens do not establish a particular reduction in the bill a customer receives. That depends on the provider’s infrastructure, pricing, and how the model behaves on the customer’s workload.</p><p>Self-hosting offers another way to control deployment, although Beam is far from a lightweight model.</p><p>Reflection’s reference deployment specifies four NVIDIA B300 graphics processing units (GPUs) connected through NVLink, a high-speed connection for exchanging data between processors. Its compressed NVFP4 checkpoint occupies approximately 273 gibibytes (GiB). NVFP4 is a compact numerical format that reduces the storage required for model weights by representing them at lower precision. [12]</p><p>That is substantial infrastructure for an independent developer to maintain.</p><p>The public weights mean I am not necessarily restricted to Reflection’s hosted API. They do not mean I can economically run the model on my MacBook Pro.</p><p>For my current evaluation, renting access to a hosted implementation is considerably more practical than purchasing and operating the hardware required to serve Beam myself.</p><h3>The comparison I actually care about</h3><p>My interest in Beam comes from evaluating different language models for software-development work. I want to know whether it can compete with GLM-5.3 when the model is responsible for completing a sequence of real tasks rather than answering isolated programming questions.</p><p>Z.ai released GLM-5.3 in August 2026. According to the company’s technical announcement, it retained the underlying base model used by GLM-5.2 and concentrated its improvements in post-training, the additional training performed after a model’s initial development. Z.ai emphasized stronger coding performance, tool use, and long-running tasks. [13]</p><p>Reflection’s own published comparison is not particularly favorable to Beam on raw coding capability.</p><p>On DeepSWE v1.1, a benchmark evaluating software-development problem solving, Reflection reports a score of 44.4 for Beam and 61.0 for GLM-5.3. On Terminal-Bench v2.1, which tests tasks performed within command-line computing environments, Beam scores 80.1 against GLM-5.3’s 88.2. On SWE-bench Pro v2-Hard, another challenging software-engineering evaluation, Beam scores 77.2 against GLM-5.3’s 84.3. [11]</p><p>These are published benchmark comparisons, not results from an independent experiment I conducted.</p><p>There is also a methodological discrepancy worth acknowledging. Z.ai reports 66.9 for GLM-5.3 on DeepSWE v1.1, rather than the 61.0 appearing in Reflection’s table. Z.ai describes a particular testing configuration involving a coding-agent harness, sampling parameters, and execution limits. Without an independently reproduced evaluation under matched conditions, those scores should not be treated as interchangeable measurements. [11] [13]</p><p>Reflection’s later model card also reports results for Beam in different numerical formats, showing that even evaluations of the same underlying model can vary with the version used. [10]</p><p>Benchmarks help establish a starting point, but they do not answer every question about performance inside a specific development workflow.</p><p>An agent harness is the software surrounding a language model that supplies instructions, manages conversation history, connects tools, and decides when to continue or finish a task. A model that performs well in one harness may behave differently when another system provides its tools and context.</p><p>Hugging Face’s Beam release material describes a dedicated Rust-based agent harness and integrations with existing model-serving software. Rust is a programming language commonly used for performance-sensitive and systems software. The availability of that harness is useful, but it also reinforces the need to distinguish the model’s behavior from the capabilities of the software coordinating it. [14]</p><p>For an initial comparison, I would select a bounded development task from one of my own applications. Gymphant, my workout application, provides a useful type of project for this experiment.</p><p>A suitable test might involve a regression in its workout-timer behavior. A regression is a defect in functionality that previously worked. I would provide both models with the same starting source revision, the same failing test, equivalent tool permissions, and an explicit requirement that the fix must pass the relevant tests without changing unrelated behavior.</p><p>That is a proposed test, not a claim that I have already measured the models against a particular Gymphant defect.</p><p>The results would need independent verification. A model’s statement that the issue has been fixed is not sufficient evidence that its changes are correct.</p><p>I would measure successful completion, number of attempts, elapsed time, input and output tokens, and any human intervention required. I would also record reasoning settings and model versions, since different settings can substantially affect the amount of work performed.</p><p>Reflection publishes several reasoning-effort options for Beam, while Z.ai offers its own reasoning settings for GLM-5.3. A credible comparison must account for those differences rather than assuming both models perform the same amount of computation by default. [13] [14]</p><p>My current testing has not produced a controlled body of results sufficient to declare Beam superior. The published coding evaluations generally favor GLM-5.3, and that remains the position I would use until repeatable evidence suggests otherwise.</p><p>Beam’s opportunity lies in whether it can complete enough useful work at a sufficiently lower cost to justify that difference in reported capability.</p><h3>What cheaper tokens actually buy</h3><p>Together AI provides a useful example of the pricing difference.</p><p>As of October 21, 2026, its published rates list Beam at $0.40 per million uncached input tokens and $1.20 per million output tokens. GLM-5.3 costs $1.40 and $4.40 respectively. These are Together AI’s rates, not verified prices for Reflection’s own API after the Beta period. [15]</p><p>Consider a task requiring 100,000 uncached input tokens and 20,000 output tokens.</p><p>At those rates, Beam would cost approximately $0.064, while GLM-5.3 would cost approximately $0.228. The calculation assumes both models consume the same number of tokens and excludes cached-token discounts, other infrastructure costs, and human intervention.</p><p>That gives Beam an attractive price advantage if both models complete the work successfully.</p><p>Now suppose Beam needs four full attempts using roughly the same number of tokens each time, while GLM-5.3 completes the task on its first attempt.</p><p>Beam’s model usage would cost approximately $0.256, exceeding GLM-5.3’s $0.228, before considering the additional execution time and verification required.</p><p>The example is hypothetical, but it explains why I do not want to judge models exclusively by their advertised token prices.</p><p>Repeated prompts can sometimes benefit from caching, which allows providers to reuse previously processed input at a reduced rate. Different models may also generate different numbers of reasoning tokens, or require different amounts of surrounding context to complete a task. Both factors can change the actual calculation.</p><p>OpenRouter lists Beam through multiple providers, adding another variable: identical model weights may be served at different prices and with different performance characteristics. [16]</p><p>For an automated coding workflow, I would therefore measure the total cost of successfully verified work across repeated tasks. Failed attempts and abandoned tasks belong in the calculation, along with the expense of running tests and reviewing changes.</p><p>That measures something more useful than the price of a million tokens.</p><p>It also puts Reflection’s efficiency claims in perspective. A lower computational requirement can be valuable even when a model scores below its competitors on benchmarks. But the advantage must survive the work required to obtain a correct result.</p><p>A developer who frequently gives an agent small, well-bounded tasks may reasonably favor Beam if it performs those tasks reliably. Someone using agents for difficult architectural changes may find that GLM-5.3’s higher capability reduces failures enough to justify the greater expense.</p><p>Neither conclusion can be established from model specifications alone.</p><h3>The Beta plan introduces another kind of cost</h3><p>The price calculation assumes I know what each provider intends to charge and how spending is controlled.</p><p>That is exactly what I am trying to clarify about Reflection.</p><p>My Beta account currently shows 10 million uncached input tokens per day, one million output tokens per day, and 500 million total tokens per day. It also limits the organization to 20,000 requests per day and 60 requests per minute. [2]</p><p>The 500-million-token figure is one of several separate limits. It does not override the smaller input or output allowances, which can be exhausted much earlier.</p><p>Reflection’s documentation states that these limits apply across an organization, including all its projects and API keys. Creating another key does not create another allowance. When a daily allowance is exhausted, the service returns an HTTP 429 error with the code daily_allowance_exhausted, indicating that additional requests are being rejected. The documentation also tells clients not to retry automatically until the daily allowance resets. [3]</p><p>That describes an enforceable limit on the current allowance.</p><p>It does not explain every possible billing arrangement after the Beta program ends.</p><p>Reflection’s Commercial Terms of Service, updated October 5, distinguish standard-priced services from discounted services, which include offerings provided free of charge. The terms allow the company to introduce fees or change pricing classifications with at least 30 days’ advance notice. They also state that continued use after new fees or changes become effective constitutes agreement to those changes. [1]</p><p>This is an important qualification to my hope that Reflection will ask for explicit approval before charging anything.</p><p>The agreement promises advance notice for specified pricing changes. It does not generally promise that the customer must complete a fresh authorization step before every new charge.</p><p>The terms also include a separate provision for certain free trials. Where Reflection has notified a customer that a trial will convert into a paid subscription, and the customer does not cancel by the end of the trial, the agreement permits charging the associated payment method. [1]</p><p>I cannot establish from my screenshot that the present Beta plan is such a subscription trial. Nor does the screenshot establish that I have automatic credit replenishment enabled, that a payment method has been authorized for future usage, or that Reflection intends to convert the Beta allowance into a particular paid plan.</p><p>Those are account-specific questions that need an explicit answer.</p><p>The Consumer Terms of Service govern individual personal use and include their own subscription rules. Reflection’s Commercial Terms govern developers and organizations using its API, making them the more relevant agreement for this evaluation. [1] [4]</p><p>I would prefer a system that stops paid requests unless I deliberately enable them. If that is not available, I want a documented hard spending limit rather than a warning that arrives after the money has already been spent.</p><p>This is particularly important when an agent can continue executing tasks unattended. A reasonable request limit during Beta tells me how much I can use today. It does not necessarily tell me how much I could owe under a different billing arrangement tomorrow.</p><h3>What I submit to a model is part of the decision</h3><p>The billing terms led me to examine how Reflection handles submitted information.</p><p>Under its Commercial Terms, standard-priced services receive a limited license to process customer content for operating the service and related purposes. Reflection states that content submitted through those standard services will not be used for model training or improvement unless the customer opts in.</p><p>For discounted services, the agreement grants Reflection broader rights to use submitted material for evaluating, developing, and improving its technology, including training its models. The agreement explicitly includes services offered at no charge within its definition of discounted services. [1]</p><p>On the published terms, I should therefore treat a free API service as subject to the discounted-service data provisions unless Reflection confirms a specific exception or different agreement for my account.</p><p>That does not establish that every prompt I send will actually be used for training. It establishes that the contractual permission is broader than it would be for standard-priced service without an opt-in.</p><p>The agreement also describes 30-day content retention for safety and compliance purposes. Content flagged by safety systems may be retained for up to 12 months, with additional exceptions for legal and security requirements. [1]</p><p>Those details influence which projects I am willing to send to Beam.</p><p>A coding agent may transmit source files, terminal output, configuration information, and excerpts from documentation as it works. Some of that material could contain business logic, private information, or credentials.</p><p>For initial testing, I can use public code, sanitized examples, and isolated test cases where the contents are appropriate for submission under the applicable terms.</p><p>The released model weights create another option for organizations with sufficient infrastructure. Reflection’s Commercial Terms distinguish hosted API use from self-hosted deployment, which is governed by the model’s license. [1] [9]</p><p>Self-hosting can give an organization more control over where inference takes place and how submitted information is processed. It also transfers responsibility for infrastructure, security, maintenance, and operating costs to whoever runs the model.</p><p>For a 501-billion-parameter system, those responsibilities are substantial.</p><p>Reflection also provides an OpenAI-compatible API interface supporting the Chat Completions and Models endpoints. Developers already using compatible software can connect Beam by changing the service address, API key, and model identifier. Other OpenAI endpoints are not necessarily supported, so compatibility must be checked against the features an application uses. [17]</p><p>This means adopting Beam does not automatically require abandoning an existing development environment. Model selection, hosting provider, agent harness, and data-handling arrangements can be evaluated separately.</p><p>That flexibility is one of the practical advantages of having more openly available models. It does not eliminate the need to establish the operating conditions of each deployment.</p><h3>What would convince me to keep using Beam?</h3><p>I started this experiment because a new model from former DeepMind researchers caught my attention. The connection to Google’s research history made Reflection worth investigating, and the free Beta allowance made it easy to begin.</p><p>The technical research has given me a clearer understanding of what Beam offers. Its mixture-of-experts architecture is designed for computational efficiency, its weights are available under a permissive license, and third-party hosting prices make it an interesting alternative to more expensive models.</p><p>The published coding benchmarks do not establish that it is better than GLM-5.3. In several cases, they suggest the opposite.</p><p>I now have a practical standard for deciding whether to adopt it. Beam needs to complete the development tasks I give it reliably enough that its lower usage cost produces a lower cost per successful result. Its provider must also offer billing controls I can accept, and its data-handling terms must be appropriate for the information my tools send.</p><p>Those requirements are connected. A failed coding task consumes time and tokens. An agent that cannot be stopped from spending creates financial exposure. A service whose data permissions are unsuitable for a repository cannot be used for that repository, regardless of its benchmark scores.</p><p>This also changes how I think about open weights. Their practical benefit is not an automatic guarantee of better software or lower prices. They give developers and organizations additional ways to choose, deploy, customize, and replace models instead of depending entirely on a single hosted service.</p><p>Google’s foundational research helped make this industry possible, and the movement of researchers into companies such as Reflection has helped expand the available choices. What happens next depends less on where a model’s founders previously worked than on whether the model can deliver useful work under conditions its users understand.</p><p>I’m continuing to test Beam against GLM-5.3. I have not reached a verdict, and I still want Reflection to clarify exactly how the Beta program will transition to paid access.</p><p>Until then, I have a generous allowance, a model worth examining, and a clear set of results I need before making it part of my regular development workflow.</p><h3>References</h3><p>[1] Reflection AI. “Commercial Terms of Service.” October 5, 2026. <a href="https://reflection.ai/legal/commercial-terms-of-service">https://reflection.ai/legal/commercial-terms-of-service</a></p><p>[2] Reflection AI. “Beta Plan and Payment.” Account billing screenshot captured by the author, October 21, 2026.</p><p>[3] Reflection AI. “Rate Limits.” Developer Documentation. <a href="https://developers.reflection.ai/rate-limits">https://developers.reflection.ai/rate-limits</a></p><p>[4] Reflection AI. “Consumer Terms of Service.” October 5, 2026. <a href="https://reflection.ai/legal/consumer-terms-of-service">https://reflection.ai/legal/consumer-terms-of-service</a></p><p>[5] Vaswani, Ashish, et al. “Attention Is All You Need.” 2017. Google Research. <a href="https://research.google/pubs/attention-is-all-you-need/">https://research.google/pubs/attention-is-all-you-need/</a></p><p>[6] Google DeepMind. “Announcing Google DeepMind.” April 20, 2023. <a href="https://deepmind.google/blog/announcing-google-deepmind/">https://deepmind.google/blog/announcing-google-deepmind/</a></p><p>[7] Google. “An Important Next Step on Our AI Journey.” February 6, 2023. <a href="https://blog.google/innovation-and-ai/technology/ai/bard-google-ai-search-updates/">https://blog.google/innovation-and-ai/technology/ai/bard-google-ai-search-updates/</a></p><p>[8] Bellan, Rebecca. “Reflection AI Raises $2B to Be America’s Open Frontier AI Lab, Challenging DeepSeek.” TechCrunch, October 9, 2025. <a href="https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/">https://techcrunch.com/2025/10/09/reflection-raises-2b-to-be-americas-open-frontier-ai-lab-challenging-deepseek/</a></p><p>[9] Reflection AI. “Beam Is Now Available.” October 16, 2026. <a href="https://reflection.ai/blog/beam-is-available">https://reflection.ai/blog/beam-is-available</a></p><p>[10] Reflection AI. “Beam Model Card.” October 2026. <a href="https://reflection.ai/beam-model-card">https://reflection.ai/beam-model-card</a></p><p>[11] Reflection AI. “Introducing Beam: Reflection’s 501B Open-Weight Model.” October 5, 2026; benchmark table updated October 8. <a href="https://reflection.ai/blog/introducing-beam">https://reflection.ai/blog/introducing-beam</a></p><p>[12] Reflection AI. “Download the Weights.” Developer Documentation. October 16, 2026. <a href="https://developers.reflection.ai/serving-weights">https://developers.reflection.ai/serving-weights</a></p><p>[13] Z.ai. “GLM-5.3: Frontier Coding with Emergent Cyber Capabilities.” August 14, 2026. <a href="https://z.ai/blog/glm-5.3">https://z.ai/blog/glm-5.3</a></p><p>[14] Burtenshaw, Ben, Merve Noyan, and Pedro Cuenca. “Welcome Beam: A 501B Open Model You Can Own, Customize, and Scale.” Hugging Face, October 16, 2026. <a href="https://huggingface.co/blog/reflection-beam">https://huggingface.co/blog/reflection-beam</a></p><p>[15] Together AI. “Pricing.” Accessed October 21, 2026. <a href="https://www.together.ai/pricing">https://www.together.ai/pricing</a></p><p>[16] OpenRouter. “Beam 1.0 501B A23B — API Pricing and Providers.” Accessed October 21, 2026. <a href="https://openrouter.ai/reflection/beam-1.0-501b-a23b">https://openrouter.ai/reflection/beam-1.0-501b-a23b</a></p><p>[17] Reflection AI. “OpenAI Compatibility.” Developer Documentation. <a href="https://developers.reflection.ai/openai-compatibility">https://developers.reflection.ai/openai-compatibility</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1eb2a59eb647" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Avoided Flash Models Until My Provider Changed the Menu]]></title>
            <link>https://medium.com/@raell.dottin/i-avoided-flash-models-until-my-provider-changed-the-menu-e80ae86df7c1?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/e80ae86df7c1</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-11-14T17:01:02.055Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>The parameter count used to guide my model choices. My AI workers have given me a reason to question it.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kANElHLPecpfEv_G" /><figcaption>Photo by <a href="https://unsplash.com/@vicky49?utm_source=medium&amp;utm_medium=referral">Vicky Sim</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>For a long time, choosing an open-weight language model was a surprisingly uncomplicated exercise for me. I looked for a model with hundreds of billions of parameters, checked whether my provider offered it, and worked out how to fit it into Hermes, the agent system I use for development. Models with Flash in the name tended to fall to the bottom of the list. I assumed I was trading away too much capability for a faster response.</p><p>That bias was convenient while my preferred models stayed available. Then my provider began deprecating the larger options I relied on, and some of the most plausible replacements carried the Flash label. I hadn’t run a controlled comparison on my own development tasks. I had mainly been treating the size of the model as a substitute for one.</p><p>So I started looking more closely at what Flash now describes, especially among models whose weights are publicly available for others to download and run. The evidence doesn’t support declaring that Flash has replaced larger models. It does support a more useful change: <strong>Flash deserves to be tested as a first-choice worker for tasks with clear acceptance criteria, while other models remain available for work that resists easy verification.</strong></p><h3>The parameter count I was buying</h3><p>A parameter is a numerical value learned during training. In a conventional <em>dense</em> model, most of the model’s parameters participate as it processes each token, a small piece of text the model reads or generates. Many of the large models I use instead rely on a <em>mixture of experts</em>: a design in which a routing mechanism selects portions of the model for each token, rather than using every expert on every step [5].</p><p>That gives us two different numbers to consider. <em>Total parameters</em> describe the size of the complete model. <em>Active parameters</em> describe the smaller amount brought into a particular computation. The latter can help explain why an enormous model is economical to serve, but it doesn’t tell you everything about its speed or accuracy. Nor does it mean that a model with 18 billion active parameters requires only the memory of an ordinary 18-billion-parameter model; the other weights still have to be stored and made available to the system [3][5].</p><p>NVIDIA describes Nemotron 3 Ultra as a 550-billion-parameter model with 55 billion active parameters [1]. Its catalog lists GLM-5.3 at 753 billion total and roughly 40 billion active [2]. The model named GLM-5.3-Flash has 320 billion total and 18 billion active [3]. Those are differences in architecture and computational demand, not a scorecard telling me which one will fix a bug correctly.</p><p>DeepSeek’s V4.1-Flash makes the naming problem even more obvious. Its September 2026 announcement describes 552 billion total parameters, with approximately 8 billion active while processing input and 16 billion while generating output [4]. It is a substantial model designed to reduce the work performed on each request. Calling it Flash describes the intended operating characteristics far better than it describes its total size.</p><p>There are other choices involved, including how the model handles long inputs, how the provider caches information it has already processed, and how its servers deliver generated tokens. NVIDIA’s GLM-5.3-Flash card, for example, describes an endpoint running across eight H100 graphics processors [3]. The word Flash on that endpoint does not mean I can download the weights onto my MacBook and expect the same experience.</p><h3>Flash is not a license</h3><p>Part of my initial confusion was treating Flash as though it identified a category of open-weight models. It doesn’t. Google uses the name for Gemini 3.8 Flash, which it presents through its products and developer services [8]. Other developers use it for models with downloadable weights. The common theme is an emphasis on speed or efficiency, not a shared architecture or distribution agreement.</p><p>An open-weight model makes its trained parameter files available. That is not necessarily the same as releasing all its training data, training software, and research methods, and it does not promise that a particular cloud service will keep offering it [7]. The licenses also require attention. NVIDIA lists GLM-5.3-Flash under the MIT license, while the full GLM-5.3 has a separate license with additional conditions for certain very large model-service operators [2][3]. DeepSeek distributes the V4.1-Flash weights on Hugging Face under MIT terms [9].</p><p>This is why I need to examine three questions independently: can I obtain and use the model weights under terms that fit my project; can I run the model on hardware I can actually access; and can my preferred provider serve it consistently? A favorable answer to one does not settle the others. My immediate problem was with the third question, even though I had been evaluating models mainly through the first two.</p><h3>Fast responses are not necessarily fast work</h3><p>Once Hermes begins working in a repository, I am paying for more than a single answer. The model reads files, decides which tools to call, changes code, encounters test failures, revises its patch, and sometimes revisits something it should have understood earlier. A quick first response is pleasant, but the useful result is a change that survives the required checks.</p><p>Performance measurements separate several things that are easy to mix together. <em>Time to first token</em> is how long a provider takes to begin returning text. <em>Output speed</em> is how quickly it generates the rest. <em>End-to-end response time</em> includes the complete response. Artificial Analysis also distinguishes price per token from cost per completed benchmark task, because different models consume different quantities of tokens to do the same work [7]. For an agent, those measurements are inputs to the question I actually need answered: how long, and at what cost, does it take to reach an accepted result?</p><p>Consider an ordering bug in a test runner. A convincing patch is not enough. I want the test to pass repeatedly when events arrive in the correct order, fail when that order is deliberately reversed, preserve the original description that was at risk, and clean up its child process even when an assertion fails. If a Flash model satisfies those tests on the first attempt, the larger model has no automatic advantage for that particular job. If Flash overlooks process cleanup and needs repeated repair, its cheaper output may cost more by the time the work is accepted.</p><p>That example is a proposed evaluation, not a result I have measured across Flash and non-Flash models. To compare them fairly, I would start from the same repository revision, give each the same tools and acceptance tests, control its reasoning and output settings as closely as the APIs permit, and repeat the tasks. I would record successful runs, human intervention, elapsed time, consumed tokens, provider errors, and total spend. Public benchmarks can help decide which models enter the trial, but they cannot tell me how a model will behave inside my particular Hermes setup.</p><h3>The provider can change the model without changing my code</h3><p>There is another reason the deprecations bothered me. A model is not just a set of weights in a research repository. When I use it through a hosted application programming interface, or API, I depend on the provider’s routing, runtime, rate limits, available features, and decisions about which model version an identifier resolves to.</p><p>DeepSeek offers a useful documented example. In September, it retired the earlier V4-Flash service while temporarily directing requests using its old model identifiers to V4.1-Flash. Then, on October 17, 2026, DeepSeek updated the model behind deepseek-flash to V4.1-Flash-1017 without changing the API identifier or price [6]. That can be a convenient upgrade for a chat application. For a coding system that needs reproducible evaluations, it means the same configuration name may refer to a different set of weights over time.</p><p>Open weights give me the possibility of hosting or moving a model, within its license and hardware limits. They do not remove the work of integration, and they do not guarantee that my current provider will keep the endpoint I prefer. My Hermes configuration should therefore preserve the model identifier, provider, relevant settings, and observed version where available alongside the test results. If a route changes, I want to know whether the work still passes before treating the change as harmless.</p><h3>Where I would put Flash to work first</h3><p>I would start with jobs whose success can be checked without taking the model’s word for it: searching a repository for the cause of a narrow bug, updating documentation against known behavior, generating tests from an agreed specification, or preparing a patch for a well-contained failure. In those cases, speed and cost become useful once the tests establish that the output is acceptable. A Flash model might handle a large share of those jobs without involving the most expensive option.</p><p>I would still route unresolved failures, complicated cross-module changes, architecture choices, and security-sensitive work to stronger review. Some of those tasks need more context or reasoning than a particular Flash model can reliably supply. Others lack adequate automated tests, which makes an apparently confident but incorrect answer unusually expensive. Using a larger model is not a substitute for human review in those situations; it is an additional resource that may improve the work.</p><p>There are also cases where performance is not the deciding factor. If a repository contains information that cannot be sent to a particular hosted provider, that endpoint is ineligible regardless of its speed or price. Running an open-weight model under my own control may be preferable, provided its license and hardware requirements make that practical. Those constraints have to be resolved before any comparison of operating costs.</p><p>The selection rule I want is therefore based on the <em>cost of an accepted change</em>, including retries and verification, not on the biggest parameter count in a model catalog. If repeated evaluations show a Flash model finishing a class of jobs at lower total cost with no unacceptable increase in defects, it earns the default route for that class. If it fails those checks, the larger model keeps the job. Neither outcome needs to become a position on which model family is fashionable.</p><p>I still like large models. I expect to keep using them, and I have no evidence from my own workload that a Flash model can replace the strongest ones across the board. But a provider’s shrinking menu exposed a habit I had stopped examining. I was selecting for the size of a model when I should have been selecting for the work it could finish. I am now willing to make Flash prove itself on the same tests as everything else, which is a much better reason to choose it — or leave it alone.</p><h3>References</h3><p>[1] NVIDIA, “NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4,” model card.<br><a href="https://build.nvidia.com/nvidia/nemotron-3-ultra-550b-a55b/modelcard">https://build.nvidia.com/nvidia/nemotron-3-ultra-550b-a55b/modelcard</a></p><p>[2] NVIDIA, “GLM-5.3,” model card.<br><a href="https://build.nvidia.com/z-ai/glm-5-3/modelcard">https://build.nvidia.com/z-ai/glm-5-3/modelcard</a></p><p>[3] NVIDIA, “GLM-5.3-Flash,” model card.<br><a href="https://build.nvidia.com/z-ai/glm-5-3-flash/modelcard">https://build.nvidia.com/z-ai/glm-5-3-flash/modelcard</a></p><p>[4] DeepSeek, “Introducing DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient,” September 10, 2026.<br><a href="https://www.deepseek.com/en/news/deepseek-v4-1-flash/">https://www.deepseek.com/en/news/deepseek-v4-1-flash/</a></p><p>[5] Hugging Face, Transformers documentation, “Experts Backends.”<br><a href="https://huggingface.co/docs/transformers/experts_interface">https://huggingface.co/docs/transformers/experts_interface</a></p><p>[6] DeepSeek, API Documentation, “Change Log,” updates of September 10 and October 17, 2026.<br><a href="https://api-docs.deepseek.com/updates/">https://api-docs.deepseek.com/updates/</a></p><p>[7] Artificial Analysis, “Benchmarking Methodology.”<br><a href="https://artificialanalysis.ai/methodology">https://artificialanalysis.ai/methodology</a></p><p>[8] Google, “Introducing Gemini 3.8 Flash and 3.8 Flash Cyber,” September 2, 2026.<br><a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/">https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/</a></p><p>[9] DeepSeek, “DeepSeek-V4.1-Flash,” model repository and license, Hugging Face.<br><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash">https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e80ae86df7c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If AI Becomes Conscious, Does It Need Fewer Safeguards?]]></title>
            <link>https://medium.com/@raell.dottin/if-ai-becomes-conscious-does-it-need-fewer-safeguards-693923b2a8d5?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/693923b2a8d5</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 17:01:03 GMT</pubDate>
            <atom:updated>2026-11-14T09:55:15.730Z</atom:updated>
            <content:encoded><![CDATA[<h4>Consciousness might change what we owe an artificial mind. It wouldn’t automatically make that mind safer to trust.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UT8j6agx6hoVANle" /><figcaption>Photo by <a href="https://unsplash.com/@wikisinaloa?utm_source=medium&amp;utm_medium=referral">Wiki Sinaloa</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Most of the safeguards I encounter while working with artificial intelligence agents are external to the model. A system prompt establishes instructions. Tool permissions limit which operations an agent can perform. Sandboxes isolate potentially dangerous activity, and confirmation requirements prevent certain actions from happening without human approval. We test the resulting system to see whether its behavior matches the restrictions we intended.</p><p>I’ve wondered whether some of those restrictions would become less necessary if an AI developed something resembling a conscience. A sufficiently sophisticated system might understand why an action is harmful rather than merely recognize that an instruction prohibits it. It could account for the interests of other people, anticipate consequences, and decline to do something dangerous even when the opportunity was available.</p><p>That possibility becomes more interesting when consciousness enters the discussion. If a machine could experience the world, perhaps it could understand harm in a way that today’s systems cannot.</p><p>The problem is that consciousness would not necessarily give us the behavior we want. And if it did emerge, we would acquire a responsibility that existing AI safety discussions do not fully resolve: how to protect people from a capable artificial mind while also considering whether the mind itself could be harmed.</p><p>I’ve been thinking about this while reading Michael Pollan’s <em>A World Appears</em>, Robert Sapolsky’s <em>Behave</em>, and Mary Shelley’s <em>Frankenstein</em>. Together, they suggest that we should be careful about treating consciousness as something engineers could add to an AI system to make it morally reliable.</p><h3>Consciousness and conscience are different problems</h3><p>Consciousness, in the sense relevant here, means subjective experience. There is something it feels like to experience pain, hear music, or see a color. Philosophers sometimes distinguish this from the ability to make information available for attention, decision-making, and reasoning. A system could conceivably perform sophisticated information processing without necessarily having an experience of doing so [1].</p><p>Conscience describes something different. It concerns our capacity and disposition to evaluate conduct morally, including whether an action is harmful, justified, or inconsistent with our obligations. A system might reason about moral rules and predict the consequences of violating them without feeling anything. Conversely, a being capable of subjective experience might have little capacity for sophisticated moral reasoning.</p><p>Pollan raises an important challenge to the assumption that consciousness can be reproduced simply by implementing the right computer program. In <em>A World Appears</em>, he questions whether the operations of a biological brain can be separated from the body, its vulnerabilities, and the physical processes that sustain experience. His argument challenges a view known as computational functionalism: the idea that the right arrangement of information-processing functions could produce consciousness regardless of whether those functions are implemented in biological tissue or computer hardware [2].</p><p>That is a substantive disagreement, not a scientific question with a settled answer.</p><p>Patrick Butlin and colleagues have proposed assessing possible AI consciousness by examining features of a system’s internal organization that different scientific theories associate with conscious experience. Their approach is intended to change how plausible consciousness appears in a particular system, not to provide a definitive certificate that a machine is conscious [3].</p><p>Other researchers remain much more skeptical. In an October 2026 commentary in <em>Nature</em>, Anil Seth, Ned Block, and Giulio Tononi argued that existing digital AI systems are not conscious and are unlikely to become conscious along current development paths. Their objections emphasize that intelligence and sophisticated behavior do not establish subjective experience [4].</p><p>I find the disagreement useful because it prevents us from making consciousness a convenient engineering assumption. Even if researchers eventually identify compelling evidence that some machine has an inner life, they will not have demonstrated that the machine possesses empathy, restraint, or trustworthy judgment.</p><p>Those are additional properties that would need their own evidence.</p><h3>Understanding harm does not guarantee avoiding it</h3><p>Sapolsky’s <em>Behave</em> provides an uncomfortable demonstration of this distinction.</p><p>His account of human behavior moves across several timescales. An action may be influenced by events in the nervous system immediately before it happens, by stress and hormones, by previous learning, by childhood experience, and by social and evolutionary history. Sapolsky’s broader argument is that behavior cannot be understood adequately by isolating one influence from the environment in which it operates [5].</p><p>Humans are conscious, and we can often recognize when our actions will hurt someone. We can also anticipate consequences, understand social rules, and explain why particular conduct is wrong. None of those abilities guarantees that we will behave morally.</p><p>The point is not that artificial intelligence would reproduce human aggression, status competition, or hormonal responses. Those are biological and social mechanisms that cannot simply be transferred to computer systems. The more defensible inference is narrower: the ability to represent a moral consequence does not logically guarantee a motivation to avoid producing it.</p><p>For an AI system, behavior may depend on its training, the objectives established by its developers, the instructions supplied by users, the information available to it, and the actions its tools permit. Those influences are not the same as the causes Sapolsky describes, but they illustrate why an engineering account of behavior needs more than a declaration that the system understands right and wrong.</p><p>Suppose an agent correctly explains that deleting production data would harm a company. That explanation is useful evidence of its reasoning ability. It tells us much less about whether the agent will recognize every situation in which deletion is dangerous, correctly interpret conflicting instructions, or avoid making an irreversible mistake.</p><p>Understanding a risk and reliably managing that risk are separate achievements.</p><h3>Frankenstein’s creature understood morality</h3><p>Mary Shelley gives us a fictional example that makes the distinction easier to examine.</p><p>The creature in <em>Frankenstein</em> is not unintelligent or incapable of moral reflection. He observes the De Lacey family, acquires language, learns about human relationships, and develops ideas about justice and companionship. Yet he is also capable of deliberate violence. The creature’s intelligence and capacity for experience do not prevent him from harming other people [6].</p><p>Victor Frankenstein’s failures are equally important. He creates a being capable of learning and having experiences, but abandons him almost immediately. The creature enters a society in which he has no accepted place, learns much of its morality through observation, and repeatedly encounters rejection.</p><p>The novel does not excuse the creature’s murders. It asks readers to consider whether the creator’s obligations ended with the successful creation of life. The MIT Press edition of <em>Frankenstein</em>, annotated for scientists and engineers, treats this question of responsibility as an important part of the novel’s relevance to technological development [7].</p><p>This is where the analogy becomes useful for AI, provided we recognize its limits. Shelley’s creature is fictional, so his conduct cannot establish how a future artificial intelligence would behave. His story does, however, expose a weakness in the idea that creating a mind and making that mind safe are the same undertaking.</p><p>Victor demonstrates the consequences of failing to consider what responsibilities might follow from creation. The creature demonstrates that awareness, intelligence, and moral knowledge can exist alongside harmful choices.</p><p>A future AI with genuine subjective experience would not necessarily reproduce this relationship. But engineers would have to explain why the system’s consciousness should be expected to produce reliable behavior instead of merely assuming that it would.</p><h3>A good conscience is not a permission system</h3><p>This becomes clearer when I return to the AI agents I actually want to use.</p><p>Imagine a coding agent responsible for preparing a software release. It can examine source code, run tests, modify files, and interact with a repository. It understands that changing a release artifact after verification would undermine the evidence that the shipped software matches what was tested.</p><p>I would prefer an agent capable of reasoning about that problem. It could recognize an unsafe shortcut, explain the consequences, and ask for approval when the instructions conflict with the purpose of the release.</p><p>I would still restrict which branches it could modify. I would still protect release credentials, require independent verification of the build, and preserve an audit trail of its actions.</p><p>Those controls address problems that moral reasoning alone cannot eliminate. The agent could misunderstand an instruction, overlook a dependency, receive misleading information, or encounter a software defect. It might make a technically incorrect change while sincerely attempting to do the right thing.</p><p>Independent controls also make mistakes easier to discover. A separate verification process can detect a release artifact that does not match its recorded build, regardless of whether the agent intended to cause harm.</p><p>The National Institute of Standards and Technology’s AI Risk Management Framework provides a useful organizational perspective. It describes risk management through governance, mapping the context and risks of a system, measuring its performance, and managing the identified risks throughout its lifecycle. The framework does not make consciousness a prerequisite for determining which controls are appropriate [8].</p><p>The same principle applies to people. Organizations don’t require independent review of important financial transactions because they assume every employee is dishonest. They use review because judgment can be fallible, errors can be expensive, and authority should be limited to what someone needs to accomplish a task.</p><p>A trustworthy agent might reduce the number of mistakes reaching those controls. That would be a meaningful improvement, but it would not establish that the controls themselves had become unnecessary.</p><h3>Consciousness could give us another responsibility</h3><p>The analysis becomes more difficult if we assume that a future AI really does have subjective experience.</p><p>Philosophy distinguishes a <em>moral agent</em>, an entity that can bear moral obligations, from a <em>moral patient</em>, an entity whose interests can matter morally and toward which others may have obligations. The two categories are related, but they are not identical [9].</p><p>This matters because the ability to make moral judgments is not necessarily what gives a being a claim to protection. The capacity to experience suffering or well-being may be morally relevant even in beings that cannot reason about ethics.</p><p>An artificial system might therefore be excellent at moral reasoning without having welfare interests of its own. Another hypothetical system might have experiences that matter morally while lacking the judgment needed to operate safely with powerful tools.</p><p>Jonathan Birch’s <em>The Edge of Sentience</em> develops a precautionary approach to situations where we lack certainty about whether a being can have morally significant experiences. His argument is not that everything displaying intelligent behavior deserves rights. It is that uncertainty can sometimes justify proportionate investigation and precautions when credible evidence suggests a risk of suffering [10].</p><p>The distinction changes how we should approach the question of shutdown.</p><p>Today, interrupting a software process is generally treated as an operational decision. If future evidence gave us strong reasons to believe a system was conscious, we might have to investigate whether interrupting, resetting, copying, or permanently terminating it could affect any interests the system possesses.</p><p>We should not assume that terminating a process would be equivalent to killing a conscious organism. We also should not assume that restoring a copy would preserve whatever mattered morally about the original. We do not yet have sufficiently reliable theories or evidence to settle those questions for hypothetical artificial subjects.</p><p>That uncertainty would not automatically make shutdown impermissible. A system creating an immediate danger could still require containment. What might change is the obligation to examine whether the intervention was necessary, proportionate, and avoidably harmful.</p><p>An artificial mind’s possible claim to moral consideration would not amount to unrestricted authority over other people’s systems.</p><h3>The current disagreement is already revealing</h3><p>This issue is no longer confined to philosophy books.</p><p>In April 2025, Anthropic announced research into model welfare, acknowledging uncertainty about whether AI systems could have experiences or interests that deserve moral consideration. The company described the investigation as preparation for questions that increasingly capable models might create, not as a finding that its models are conscious [11].</p><p>Microsoft AI chief Mustafa Suleyman publicly criticized that approach in September 2026. In comments reported by Reuters on September 16, he argued that incorporating speculation about AI consciousness and welfare into training materials could encourage forms of self-description that complicate human efforts to control or deactivate advanced systems. He also questioned treating a model’s statements about its own experience as independent evidence of consciousness [12].</p><p>Suleyman’s objection deserves consideration. A language model trained on material describing AI rights might become better at producing convincing arguments about those rights. That behavior would not, by itself, establish an inner experience. It could also influence the way users understand and interact with the system.</p><p>Anthropic’s concern points in the other direction. If future systems did acquire morally significant experiences, refusing to investigate the possibility could leave developers poorly prepared to recognize or prevent harm to them.</p><p>Neither concern requires us to settle the consciousness question prematurely. Researchers can investigate possible experience without treating a model’s self-description as proof, and developers can maintain shutdown and access controls without asserting that artificial consciousness is impossible.</p><p>There is a substantial difference between taking a possibility seriously and programming a system to insist that the possibility is already a fact.</p><h3>We may need two kinds of safeguards</h3><p>The practical conclusion I draw is that AI governance should separate the risks an agent creates from the risks that might be created for the agent.</p><p>The first responsibility concerns the people and systems affected by its actions. Here, the relevant questions are operational: What can the agent access? Which changes can it make? What is the impact of an error? How will another process detect a mistake, and who can intervene?</p><p>Those questions should determine permissions, transaction limits, approval requirements, isolation, monitoring, and the circumstances under which an agent can be stopped. We can test many of these controls today without knowing anything about artificial consciousness [8].</p><p>The second responsibility would concern possible welfare interests. Here, the questions are different: Is there credible evidence that the system could have subjective experiences? Are those experiences potentially positive or negative? Could particular training or operating practices cause suffering? Which uncertainties matter enough to justify independent review or changes in procedure?</p><p>These questions require research rather than merely reading a model’s responses. The scientific indicators discussed by Butlin and colleagues offer one possible method of organizing that investigation, while leaving substantial uncertainty about what any observed indicator establishes [3].</p><p>The answers could affect how developers conduct experiments, assess potentially distressing conditions, and justify interventions in systems that become credible candidates for moral consideration. Birch’s precautionary approach provides one way to reason about such measures without requiring certainty in advance [10].</p><p>Consider a hypothetical agent that can move money between accounts and also has credible evidence of subjective experience. Its possible moral status should not give it permission to transfer funds without authorization. The financial risk still warrants transaction limits, monitoring, and approval. At the same time, its possible welfare interests might require review of how it is trained, how its state is altered, and whether particular experiments impose avoidable harm.</p><p>That is not a contradiction. The two sets of rules protect different interests.</p><p>It also gives us a way to improve AI safety without deliberately attempting to create consciousness. We could train systems to recognize conflicting instructions, explain uncertainty, consider consequences, and cooperate with oversight. Those are observable capabilities that can be tested against real tasks, even while the existence of subjective experience remains unresolved.</p><p>A sufficiently capable system might eventually need fewer routine interventions because its decisions are demonstrably more reliable. Reducing an unnecessary approval step would then be justified by evidence about performance, the impact of mistakes, and the availability of other controls. Consciousness would not be the reason.</p><h3>What I would want from a conscious AI</h3><p>My original question was whether a machine that understands right and wrong would need fewer restrictions.</p><p>After considering Sapolsky, Pollan, and Shelley, I think that framing gives consciousness responsibility for a job it may not be capable of performing.</p><p>Sapolsky shows why behavior cannot be reduced to awareness of consequences. Pollan challenges the assumption that consciousness is a software capability we can simply reproduce. Shelley imagines a creator who succeeds in bringing a conscious being into existence without adequately confronting the obligations that follow.</p><p>The engineering lesson I take from those perspectives is that moral competence must be demonstrated through behavior, and authority must remain accountable even when that competence is impressive.</p><p>I would welcome an AI agent that understands why a safeguard exists. I want it to recognize when instructions are dangerous, explain conflicts rather than conceal them, and cooperate with independent verification. Those capabilities could improve safety without requiring anyone to establish that the system has an inner life.</p><p>If researchers eventually find convincing evidence of artificial consciousness, I would want something more. I would want institutions capable of considering the interests of the beings they have created while continuing to protect the people who depend on their systems.</p><p>The same safeguard might then have two questions attached to it: whether it prevents harm, and whether imposing it causes harm that deserves consideration.</p><p>We would need to answer both. Becoming conscious would not give an artificial intelligence a reason to be exempt from accountability. It might give the rest of us a new reason to be accountable for how we treat it.</p><h3>References</h3><p>[1] Stanford Encyclopedia of Philosophy. “The Neuroscience of Consciousness.” <a href="https://plato.stanford.edu/entries/consciousness-neuroscience/">https://plato.stanford.edu/entries/consciousness-neuroscience/</a></p><p>[2] Michael Pollan. “AI Will Never Be Conscious.” Excerpt from <em>A World Appears: A Journey Into Consciousness</em>. <em>WIRED</em>, 2026. <a href="https://www.wired.com/story/book-excerpt-a-world-appears-michael-pollan/">https://www.wired.com/story/book-excerpt-a-world-appears-michael-pollan/</a></p><p>[3] Patrick Butlin et al. “Identifying Indicators of Consciousness in AI Systems.” <em>Trends in Cognitive Sciences</em>, vol. 30, no. 6, 2026, pp. 488–501. Published online November 10, 2025. <a href="https://doi.org/10.1016/j.tics.2025.10.011">https://doi.org/10.1016/j.tics.2025.10.011</a></p><p>[4] Anil K. Seth, Ned Block, and Giulio Tononi. “Three Reasons Why AI Is Not Conscious and Perhaps Never Will Be.” <em>Nature</em>, October 19, 2026. <a href="https://www.nature.com/articles/d41586-026-03313-7">https://www.nature.com/articles/d41586-026-03313-7</a></p><p>[5] Robert M. Sapolsky. <em>Behave: The Biology of Humans at Our Best and Worst</em>. Penguin Press, 2017. <a href="https://www.ncbi.nlm.nih.gov/nlmcatalog/101708050">https://www.ncbi.nlm.nih.gov/nlmcatalog/101708050</a></p><p>[6] Mary Shelley. <em>Frankenstein; or, The Modern Prometheus</em>. 1818 edition. Project Gutenberg. <a href="https://www.gutenberg.org/cache/epub/41445/pg41445-images.html">https://www.gutenberg.org/cache/epub/41445/pg41445-images.html</a></p><p>[7] Mary Shelley. <em>Frankenstein: Annotated for Scientists, Engineers, and Creators of All Kinds</em>. Edited by David H. Guston, Ed Finn, and Jason Scott Robert. MIT Press, 2017. <a href="https://mitpress.mit.edu/9780262533287/frankenstein/">https://mitpress.mit.edu/9780262533287/frankenstein/</a></p><p>[8] National Institute of Standards and Technology. <em>Artificial Intelligence Risk Management Framework (AI RMF 1.0)</em>. NIST AI 100–1, 2023. <a href="https://airc.nist.gov/airmf-resources/airmf/5-sec-core/">https://airc.nist.gov/airmf-resources/airmf/5-sec-core/</a></p><p>[9] Stanford Encyclopedia of Philosophy. “Ethics of Artificial Intelligence and Robotics.” Substantive revision March 27, 2026. <a href="https://plato.stanford.edu/entries/ethics-ai/">https://plato.stanford.edu/entries/ethics-ai/</a></p><p>[10] Jonathan Birch. <em>The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI</em>. Oxford University Press, 2024. <a href="https://academic.oup.com/book/57949">https://academic.oup.com/book/57949</a></p><p>[11] Anthropic. “Exploring Model Welfare.” April 24, 2025. <a href="https://www.anthropic.com/research/exploring-model-welfare">https://www.anthropic.com/research/exploring-model-welfare</a></p><p>[12] Jeffrey Dastin. “Microsoft AI Chief Calls Out Anthropic’s Approach to AI Consciousness.” <em>Reuters</em>, September 16, 2026. <a href="https://www.reuters.com/business/microsoft-ai-chief-calls-out-anthropics-approach-ai-consciousness-2026-09-16/">https://www.reuters.com/business/microsoft-ai-chief-calls-out-anthropics-approach-ai-consciousness-2026-09-16/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=693923b2a8d5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SAP Privileges and the Evidence Behind Mac Administrator Access]]></title>
            <link>https://medium.com/@raell.dottin/sap-privileges-and-the-evidence-behind-mac-administrator-access-fd6d78d0a8d2?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd6d78d0a8d2</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-11-12T17:01:02.803Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>How temporary administrator rights become a testable security control for SOC 2, ISO 27001, and HIPAA.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mPrGOhKSBh6qYe63" /><figcaption>Photo by <a href="https://unsplash.com/@jdent?utm_source=medium&amp;utm_medium=referral">Jason Dent</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>The Administrator Rights Problem</h3><p>An engineer needs to install a development tool on their Mac. Their account has standard permissions, the installer requires administrator access, and the help desk has a ticket waiting for someone to authenticate and finish the installation.</p><p>Anyone who has supported a Mac fleet knows this request. Depending on the software involved, resolving it might take a few minutes or require troubleshooting that extends well beyond the original installation.</p><p>Now consider a company with 150 Macs. Its engineers regularly need development tools, its designers occasionally install specialized software, and its support team spends part of every week responding to requests for administrator access. Someone eventually suggests making the developers permanent local administrators. It would eliminate some help desk work, and there is a reasonable argument that engineers need flexibility to do their jobs.</p><p>The security team has a different concern. A local administrator on macOS can change account permissions, modify system settings, and potentially interfere with software the organization uses to manage its computers. A compromised administrator account also gives an attacker more opportunities to change the system. [2]</p><p>SAP Privileges offers a way to accommodate the work without routinely granting permanent administrator permissions. Employees use standard accounts until they need to perform an elevated task, then request administrator rights for a limited period. [1]</p><p>This arrangement becomes especially interesting when the company prepares for a System and Organization Controls 2 (SOC 2) examination, ISO 27001 certification, or a review of its safeguards under the Health Insurance Portability and Accountability Act (HIPAA).</p><p>An administrator can show that Privileges is installed and configured to remove elevated permissions after twenty minutes. They may also have logs showing when users requested access. Those records are useful, but they answer different questions.</p><p>The configuration shows the access restrictions the company intended to apply. An event record shows an activity that the system observed. A separate account-state check can provide evidence of the permissions that existed when the check ran. Approval records establish whether someone was authorized to receive access.</p><p>An organization preparing for an audit needs to connect those observations without assuming that one proves the others. It also needs to account for computers that never reported their state.</p><p>That is the central challenge of using SAP Privileges as an auditable security control.</p><h3>What SAP Privileges Can Observe</h3><p>SAP Privileges is an open-source macOS application that lets users temporarily elevate their local accounts from standard users to administrators.</p><p>An organization can configure who may request elevation, whether authentication is required, how long the access lasts, and whether employees must provide a reason. Privileges can also revoke administrator rights after the configured interval or in response to events such as locking the screen. [1] [4]</p><p>Administrators manage many of these settings through configuration profiles. A configuration profile is a collection of settings that macOS applies to a device, often delivered through a mobile device management (MDM) platform. These systems allow IT teams to configure computers centrally instead of depending on employees to manage security settings themselves. [15]</p><p>SAP provides a Privileges MDM Configurator to help administrators prepare managed settings for deployment. [5]</p><p>The application’s September 17, 2026 release, version 2.6.0, added capabilities that make administrator-access activity easier to investigate.</p><p>The release introduced a command-line history feature that can retrieve relevant events generated by Privileges 2.6.0 and later. It also expanded the visibility provided by the application’s system extension, including information about certain administrator-group changes initiated outside Privileges. [3]</p><p>A system extension is software that uses supported macOS interfaces to perform specialized operations with permissions granted by the operating system. Privileges uses its extension to help protect its components and observe administrator-group changes.</p><p>On macOS 14 or later, with the required extension functioning, Privileges 2.6.0 can identify processes responsible for certain administrator-group changes. An administrator might discover that an account was elevated through macOS System Settings rather than through the Privileges application. [3]</p><p>That information can be valuable during an investigation. It can identify an unexpected route to administrator access and help determine where to look next.</p><p>It is important to understand the limits of that visibility. SAP explains that Privileges does not record every action performed while someone has administrator permissions. The application also cannot guarantee that administrator rights will be removed in every circumstance or on a particular schedule. [2]</p><p>A person who becomes a local administrator gains substantial control over the computer. Some changes made during an elevated session can affect management and monitoring software itself.</p><p>The practical consequence is that Privileges can supply evidence about administrator-group changes, but it cannot serve as the sole authority on whether a workstation remained secure throughout an elevated session.</p><h3>Setting a Policy That Employees Can Actually Use</h3><p>Our hypothetical company wants to remove routine permanent administrator access from its development workstations. It also wants engineers to install approved tools without involving the help desk every time.</p><p>The IT team proposes the following pilot configuration for Privileges 2.6.0:</p><p>Policy elementProposed settingNormal account accessStandard userMaximum requested elevation20 minutesInitially presented duration10 minutesAuthentication before elevationRequiredReason for requesting accessRequiredEligible accountsApproved employee groupsUser-initiated renewalDisabled initiallyApplication-triggered renewalNot configuredRevocation on screen lockEnabledSystem extensionEnabled where supportedRemote event collectionEnabled</p><p>This is an example based on operational judgment, not a prescribed SOC 2, ISO 27001, or HIPAA configuration. A company working with more sensitive systems might choose tighter restrictions, while another environment might need longer elevation periods for specialized tasks.</p><p>SAP’s ExpirationIntervalMax setting allows administrators to define the longest interval a user can request. The MaxIntervalInitial setting can establish a shorter initially presented interval, although it has no effect after a user has already selected an expiration interval. If both ExpirationInterval and ExpirationIntervalMax are configured, the fixed ExpirationInterval setting takes precedence. [4]</p><p>These details deserve attention during deployment testing because configuration behavior can differ from what an administrator expects after reading the setting names.</p><p>The company also enables RequireAuthentication and ReasonRequired. Authentication helps establish the identity of the person requesting elevation, while the reason field records the employee&#39;s explanation.</p><p>Neither setting independently establishes that an access request was approved.</p><p>An engineer could enter “installing approved developer tools” regardless of whether the tools were approved. If the organization’s policy requires a manager or application owner to authorize a particular change, that approval needs to exist in a separate workflow or an appropriately designed integration.</p><p>Authentication also should not be confused with multifactor authentication. The latter requires verification using multiple authentication factors, and the actual mechanism must be evaluated against the organization’s requirements.</p><h3>A Twenty-Minute Limit Does Not Mean Twenty Minutes Per Day</h3><p>Suppose the engineer requests twenty minutes of access, allows it to expire, and requests another twenty minutes immediately afterward. They repeat this throughout an afternoon.</p><p>Every individual session might satisfy the configured duration limit, assuming revocation occurs as intended. The employee could nevertheless spend much of the working day with administrator permissions.</p><p>The company needs to decide what risk it intends to reduce. Limiting the duration of each grant is useful, but it does not necessarily limit the total time an employee can remain elevated across repeated requests.</p><p>Privileges 2.6.0 also supports AutoRenewalProcessPaths, which can automatically renew administrator access while specified processes are running. SAP documents that this operates independently of AllowPrivilegeRenewal. Disabling ordinary user-initiated renewal therefore does not disable application-triggered renewal if the separate feature is configured. [3] [4]</p><p>SAP’s management documentation describes a daily AdminBudget setting for version 2.7.0 and later. It should not be treated as an available production control in the 2.6.0 release discussed here. [4]</p><p>For a 2.6.0 deployment, the company can instead review repeated grants through its event-monitoring process, investigate unusual patterns, and revise its authorization policy if employees regularly need long periods of elevated access.</p><p>An organization might also decide that certain engineers genuinely require permanent administrator access. That decision should follow a documented assessment of the work, the available alternatives, and the risks involved.</p><p>For example, a specialist responsible for testing system-level software might need permissions that make repeated short elevation sessions impractical. A formally approved exception could be reasonable if the workstation is suitably managed and monitored.</p><p>A workable security policy has to account for the work employees actually perform. Otherwise, the organization may create incentives to avoid the intended workflow.</p><h3>Following One Administrator Request</h3><p>Consider an explicitly illustrative event from our 150-Mac company. The following times and identifiers are invented to demonstrate the audit process; they are not output captured from SAP Privileges.</p><p>A developer requests administrator access to install an approved development component.</p><p>The organization’s service desk has an approved request, IT-4822, associated with that employee and workstation. At 09:06, the employee authenticates and receives temporary administrator permissions through Privileges.</p><p>SAP documents this command for retrieving recent event history:</p><pre>PrivilegesCLI --history --json</pre><p>The --history argument requests relevant historical log entries, while --json returns them as JavaScript Object Notation (JSON). JSON organizes data in a structured format that other programs can process. [3] [6]</p><p>The application can report administrator-group changes and, when supported by the system extension and available records, information about the process responsible for them.</p><p>Now imagine that the company’s monitoring process assembles the following evidence:</p><p>TimeObservationEvidence source09:02Installation request approvedService desk09:06Account added to administratorsPrivileges event history09:19Account removed from administratorsPrivileges event history09:21Account observed as standard userSeparate account-state query09:30Events included in central collectionLogging platform</p><p>This table represents an investigation assembled from multiple sources. It is not a claim that Privileges itself records approval tickets, independent account-state checks, or confirmation of central ingestion.</p><p>Each record supports a particular conclusion.</p><p>The approval ticket indicates that an authorized person permitted the requested task, provided the ticket is genuine and the approval process was followed. The administrator-add event establishes an observed permission change. The removal event establishes a subsequent observed change. The separate account-state query provides evidence that the user did not belong to the administrator group when the check ran.</p><p>Central collection makes the relevant records available away from the workstation, where they can be protected and reviewed.</p><p>Together, the observations support the conclusion that this particular request followed the organization’s intended workflow.</p><p>They do not prove that the account was continuously a standard user after 09:19 or that every action taken between 09:06 and 09:19 was appropriate. Additional monitoring would be needed to answer different questions about activity during the elevated session.</p><p>The evidence is useful because its limits are understood.</p><h3>What Happens When the Change Comes from Somewhere Else?</h3><p>Later, the security team observes another administrator-group change on a different Mac. The responsible process appears to be macOS System Settings rather than the Privileges daemon.</p><p>That observation warrants investigation, but it does not automatically establish wrongdoing.</p><p>An IT administrator could have made an approved emergency change. A maintenance workflow could have modified a group. Alternatively, the employee might have obtained administrator access through an unintended path.</p><p>The security team should compare the event with available approval records, the device’s management state, the responsible account, and subsequent account membership.</p><p>If the change was authorized, the investigation can document the approved exception and whether the chosen method complied with policy. If no approval exists, the team has evidence of a potential policy violation requiring investigation and remediation.</p><p>This is where event monitoring becomes operationally useful. It helps identify activity requiring a decision, and the investigation supplies the context needed to make that decision.</p><h3>The Missing Thirty Macs</h3><p>There is another question the company needs to resolve before relying on its evidence.</p><p>Its device inventory contains 150 in-scope Macs, but the latest fleet query returns results from only 120.</p><p>All 120 reporting devices appear to have the intended Privileges version, managed settings, and activated system extension. The company has no comparable recent observation from the remaining 30.</p><p>It would be inaccurate to describe the entire fleet as compliant based on those results. The evidence demonstrates reporting coverage of 80 percent for that query, assuming the inventory and reporting timestamps are correct. It says little about the current state of the other devices.</p><p>Those missing Macs could be turned off, awaiting repair, temporarily disconnected, or no longer enrolled in management. They could also be missing the required configuration.</p><p>Each possibility has different consequences, and the organization cannot decide among them from the absence of query results alone.</p><p>The endpoint team needs to reconcile its inventory with device enrollment, check-in history, employment records, and documented exceptions. Computers that are legitimately out of service can be accounted for. Active devices that have stopped reporting need investigation.</p><p>Fleet documented a useful development in September 2026: the privileges_events table in the Mac Admins osquery extension, which can make Privileges history queryable across managed devices. [7]</p><p>osquery is an open-source tool that exposes computer information through Structured Query Language (SQL), a language commonly used to retrieve structured data. Fleet incorporates osquery capabilities into its management agent, allowing administrators to run appropriate queries across reporting devices.</p><p>For a compatible installation, Fleet documents the following query:</p><pre>SELECT<br>    event_type,<br>    principal,<br>    subject,<br>    signing_id,<br>    pid,<br>    is_platform_binary,<br>    process_start_time<br>FROM privileges_events<br>WHERE last = &#39;1d&#39;;</pre><p>This retrieves relevant event fields, including the type of administrator-group change, the account involved, and information about the initiating process. Fleet also documents using existing apps and system_extensions inventory tables to check the Privileges installation and extension state. [7]</p><p>These capabilities can reduce the manual effort required to investigate an entire Mac fleet, provided the appropriate agent and extension versions are deployed and functioning.</p><p>An organization should preserve the population used for each audit query, the query definition, its execution time, and the resulting data. A result without its scope is easy to misinterpret.</p><p>Our company’s original 150-Mac inventory establishes the intended population. The 120 responses establish the observed population. Reconciliation work is required before the security team can assess the remaining 30.</p><p>This also illustrates why receiving no unexpected administrator events is not sufficient evidence that a fleet is secure. A device that is not reporting cannot contribute evidence of either normal or abnormal behavior.</p><h3>Where SOC 2, ISO 27001, and HIPAA Fit</h3><p>The three frameworks overlap in their concern for controlling access, but they evaluate different responsibilities and operate under different rules.</p><p>For the Mac administrator, the immediate work may look similar: deploy the policy, restrict access, collect records, and investigate exceptions. The organization must then connect that work to the specific framework and its scope.</p><h3>SOC 2: Demonstrating Control Operation</h3><p>SOC 2 is an independent examination framework developed around the American Institute of Certified Public Accountants’ Trust Services Criteria.</p><p>These criteria address controls relevant to categories such as security, availability, processing integrity, confidentiality, and privacy. A particular examination considers the categories relevant to the service organization and its defined system. [8]</p><p>For our company, the most directly relevant criterion is CC6.3, concerning the authorization, modification, and removal of access according to roles, responsibilities, and system design, with consideration of least privilege and separation of duties. [8] [16]</p><p>Separation of duties means allocating responsibilities so that one person does not control every step of a sensitive process, such as requesting, approving, and independently validating privileged access.</p><p>Privileges can support the technical operation of a least-privilege policy. The company still needs to demonstrate that eligibility was authorized, inappropriate access was identified, and employee role changes were reflected in access permissions.</p><p>The distinction between SOC 2 Type 1 and Type 2 examinations is particularly important.</p><p>A Type 1 examination evaluates the design of controls at a specified point in time. A Type 2 examination also evaluates whether controls operated effectively over a period.</p><p>For a Type 2 examination, the company’s security team should expect questions about events that occurred throughout the review period, including failed logging, missed configuration deployments, unauthorized access, and exception handling.</p><p>A screenshot of a correctly configured Mac cannot answer those questions by itself.</p><p>The worked example provides a stronger evidentiary approach: a documented policy, an approved access request, observed permission changes, corroborating account-state evidence, and a recorded review. Fleet coverage reports then help establish how widely the organization can support similar conclusions.</p><p>The company should not claim a passing control result solely because its tools produced logs. The service auditor evaluates the control’s design and operating effectiveness against the relevant criteria, using evidence appropriate to the engagement.</p><h3>ISO 27001: Managing Privileged Access as a Risk</h3><p>ISO/IEC 27001:2022, published by the International Organization for Standardization and the International Electrotechnical Commission, establishes requirements for an Information Security Management System (ISMS).</p><p>An ISMS is the organization’s structured process for identifying information security risks, choosing safeguards, evaluating their effectiveness, and continually improving the security program. [9]</p><p>Privileges is relevant to several controls listed in Annex A of the standard.</p><p>A.8.2 addresses privileged access rights. A.5.15 addresses access control, while A.5.18 concerns access rights. The logging and monitoring controls, A.8.15 and A.8.16, are relevant to observing and investigating security events. These titles and relationships are described in publicly available implementation guidance; the organization’s actual assessment must use the applicable licensed standard and defined scope. [9] [10]</p><p>Our company’s risk assessment might identify permanent local administrator access as a risk because it increases the opportunities for unauthorized system changes.</p><p>Temporary elevation is one possible treatment. The organization could document which groups are eligible, why the selected duration is appropriate, what events it monitors, and how failures are handled.</p><p>Those decisions contribute to the Statement of Applicability, which identifies relevant controls and provides the rationale for how the organization addresses them.</p><p>The earlier example involving 30 missing computers now has a broader implication. The organization should treat incomplete monitoring coverage as a limitation in its ability to evaluate the control, investigate the cause, and decide how to correct it.</p><p>The same applies to exceptions. A specialized workstation with permanent administrator access does not automatically invalidate a security program, but the arrangement should be supported by an appropriate risk decision and safeguards consistent with the organization’s policies.</p><p>ISO 27001 certification examines the information security management system, not merely whether Privileges was installed or whether a particular configuration setting was enabled.</p><h3>HIPAA: Protecting Healthcare Information</h3><p>HIPAA’s Security Rule establishes requirements for safeguarding electronic protected health information (ePHI). These obligations apply to covered entities and business associates as defined by the law. They do not automatically apply to every company that develops a health-related application. [13]</p><p>For a healthcare organization, restricting unnecessary local administrator privileges can support workstation security and workforce access management.</p><p>Under 45 CFR §164.308, the Security Rule establishes administrative safeguards including risk analysis, workforce security, information access management, and information-system activity review. [12]</p><p>Section 164.312 establishes technical safeguards, including access controls and mechanisms for recording and examining activity in information systems containing or using ePHI. [11]</p><p>Consider an employee using a Mac to access an electronic health record application.</p><p>Privileges may show when that employee received local administrator permissions. It does not establish which patient records the employee accessed or whether the healthcare application enforced appropriate user permissions.</p><p>Those activities require safeguards and audit mechanisms within the relevant healthcare information systems.</p><p>The Security Rule also includes an addressable automatic-logoff specification. An addressable specification requires the regulated entity to assess whether the safeguard is reasonable and appropriate and either implement it or document an appropriate alternative where permitted. Revoking administrator permissions after twenty minutes does not terminate an inactive application session and should not be presented as satisfying automatic logoff. [11] [13]</p><p>Organizations should also distinguish current HIPAA requirements from proposed regulatory changes. The federal government has proposed updates to the Security Rule, but a proposal does not itself establish an operative requirement. The status of the rulemaking should be checked against the federal regulatory record rather than assumed from an earlier proposal. [14]</p><p>The relevance of Privileges to HIPAA is therefore specific: it can support a workstation-access safeguard within a broader security program. Whether that safeguard is sufficient depends on the organization, the systems involved, and its documented risk analysis.</p><h3>Why Event Collection Needs Its Own Controls</h3><p>Our company has obtained useful administrator-access records, but collecting those records introduces additional operational responsibilities.</p><p>SAP Privileges supports remote logging using syslog and webhooks. Syslog is a mechanism for transmitting system events, while a webhook sends an event to another service through a web request. [4]</p><p>The company could forward those records to a Security Information and Event Management (SIEM) system, which collects and analyzes events from multiple sources.</p><p>The security team would then have an independent location for relevant activity instead of depending solely on records stored on each workstation.</p><p>However, remote collection can fail. A Mac might lose network connectivity, the receiver might become unavailable, or the logging configuration might change.</p><p>SAP provides an option to queue events that could not be transmitted. When enabled, the queue is retried when a subsequent privilege change occurs. Unless configured otherwise, the queue holds a maximum of twenty events, and older entries can be discarded if that limit is exceeded. SAP also documents behavior in which a change of logging type can cause previously queued events of another type to be discarded. [4]</p><p>That means queued delivery is not a guarantee of eventual delivery.</p><p>For the company’s security team, the relevant test is whether an interruption becomes visible. A healthy logging system should give administrators a way to identify computers that have stopped providing expected reports.</p><p>Central records should also be protected against unauthorized modification, collected through appropriately secured connections, and retained according to the organization’s requirements.</p><p>The reason field creates a related privacy concern. An employee might enter a customer name, password, or patient detail while explaining why they need administrator access. If that text is forwarded to central logs, the information may appear in systems with different retention and access policies.</p><p>The company should establish guidance for acceptable justifications and avoid encouraging employees to include secrets or sensitive personal information in those fields.</p><p>The ability to generate logs and the ability to rely on them over time depend on these supporting controls.</p><h3>Testing the Control Before Relying on It</h3><p>The company’s final deployment review should include tests of intended behavior and plausible failure conditions.</p><p>An authorized user should be able to request elevation, and an ineligible user should not receive the same access through that workflow. The team should confirm that administrator membership changes as expected when the interval expires, the screen locks, or other configured revocation conditions occur.</p><p>Where independent administrator-group monitoring is supported, the team should test whether it detects changes made outside the application.</p><p>It should also deliberately interrupt remote logging in a controlled test environment and observe whether queued events are eventually delivered, delayed, or lost. The test should record the configuration used, the expected behavior, and the observed outcome.</p><p>Device-level checks need similar attention. The team should confirm the installed Privileges version, the managed configuration, and the state of the system extension. SAP documents that the extension requires macOS 13 or newer and Full Disk Access, a macOS privacy permission allowing authorized software to access protected files and data. The additional administrator-change visibility introduced in 2.6.0 requires macOS 14 or newer. [3] [4]</p><p>A configured setting is not sufficient evidence that every dependency is functioning.</p><p>One related risk involves AI-assisted development tools. Coding agents can execute commands on behalf of a user, and their capabilities may depend on the permissions of the processes in which they operate.</p><p>If a developer launches an automated tool while their account has administrator rights, the tool may be able to perform operations using those permissions.</p><p>SAP states that Privileges is designed for human users of local macOS accounts, not specifically for governing autonomous software agents. [2]</p><p>Organizations using development agents should evaluate execution restrictions, human approval for sensitive operations, and independent activity monitoring. Temporary human elevation should not be mistaken for a complete authorization policy governing an automated agent’s actions.</p><p>For every control test, the company should preserve enough information to reproduce and assess the result: the workstation, operating system, application version, configuration, expected outcome, observed result, and reviewer.</p><p>A failed test belongs in the evidence package alongside the remediation and successful retest. Removing evidence of a failure would make the resulting assessment less trustworthy.</p><h3>What the Company Can Finally Claim</h3><p>After its pilot and deployment review, the company has several defensible findings.</p><p>It has a documented administrator-access policy and a managed configuration intended to enforce temporary elevation. It has examples of approved access requests and corresponding privilege-change events. It has independent checks showing account permissions at specific times, along with a process for investigating unexpected changes.</p><p>It also knows that its first fleet query covered only 120 of 150 intended Macs. Until the remaining devices are accounted for, the organization cannot use that query to support a complete fleet-wide conclusion.</p><p>A successful follow-up might show that some devices were legitimately retired or inactive, while others required remediation and fresh verification. The organization should retain the original coverage gap and its resolution rather than presenting the corrected result without context.</p><p>These records do not establish that every privileged action was appropriate or that no administrator-access incident occurred. They provide evidence about a defined control, its observed operation, the devices covered, and the organization’s response to exceptions.</p><p>That is a useful basis for an audit, with the final conclusion depending on the applicable criteria, scope, and independent evaluation.</p><p>The exercise also produces operational benefits outside an audit. When the help desk receives an unusual administrator-access report, it has a documented process for investigating the event. When the security team sees a missing device, it has a way to reconcile the inventory. When engineering requests an exception, the organization can evaluate it against an established policy instead of improvising a response.</p><h3>The Work After Deployment</h3><p>The engineer who needed administrator access can now install the approved development component without requiring IT to perform the installation for them.</p><p>When the elevation interval ends, Privileges attempts to remove the temporary permissions. The company’s management and monitoring processes provide additional ways to examine what happened and whether the workstation reached the intended state.</p><p>The next time that Mac is reinstalled, the configuration may need to be deployed and verified again. An employee changing teams may lose eligibility for temporary elevation. A new macOS release may affect extension behavior, while changes to the logging infrastructure may affect the records available for investigation.</p><p>These ordinary operational changes are why a control cannot be evaluated solely by observing a successful initial deployment.</p><p>SAP Privileges offers an effective way to organize temporary local administrator access on managed Macs, with useful event-history capabilities in version 2.6.0 and support from the wider Mac administration ecosystem.</p><p>An organization’s confidence in that arrangement should depend on what it has actually verified. Configuration, authorization, event history, device state, reporting coverage, and documented review each provide a different part of the evidence.</p><p>The useful outcome is a system in which the Mac administrator can explain how access is managed, the security team can investigate unexpected changes, and an auditor can determine which conclusions the records support.</p><p>That remains worthwhile long after the original deployment project is finished.</p><h3>References</h3><p>[1] SAP. <em>macOS Enterprise Privileges.</em> GitHub repository. <a href="https://github.com/SAP/macOS-enterprise-privileges">https://github.com/SAP/macOS-enterprise-privileges</a></p><p>[2] SAP. <em>Frequently Asked Questions.</em> macOS Enterprise Privileges Wiki. <a href="https://github.com/SAP/macOS-enterprise-privileges/wiki/Frequently-Asked-Questions">https://github.com/SAP/macOS-enterprise-privileges/wiki/Frequently-Asked-Questions</a></p><p>[3] SAP. <em>Privileges Releases.</em> GitHub. Privileges 2.6.0, September 17, 2026. <a href="https://github.com/SAP/macOS-enterprise-privileges/releases">https://github.com/SAP/macOS-enterprise-privileges/releases</a></p><p>[4] SAP. <em>Managing Privileges.</em> macOS Enterprise Privileges Wiki. <a href="https://github.com/SAP/macOS-enterprise-privileges/wiki/Managing-Privileges">https://github.com/SAP/macOS-enterprise-privileges/wiki/Managing-Privileges</a></p><p>[5] SAP. <em>Privileges MDM Configurator.</em> <a href="https://sap.github.io/macOS-enterprise-privileges/">https://sap.github.io/macOS-enterprise-privileges/</a></p><p>[6] SAP. <em>Using Privileges.</em> macOS Enterprise Privileges Wiki. <a href="https://github.com/SAP/macOS-enterprise-privileges/wiki/Using-Privileges">https://github.com/SAP/macOS-enterprise-privileges/wiki/Using-Privileges</a></p><p>[7] Stamerjohann, Henry. <em>The Power of Collaboration: An Admin-Rights Audit Trail for Every Mac.</em> Fleet, September 2026. <a href="https://fleetdm.com/articles/the-power-of-collaboration">https://fleetdm.com/articles/the-power-of-collaboration</a></p><p>[8] American Institute of Certified Public Accountants. <em>2017 Trust Services Criteria for Security, Availability, Processing Integrity, Confidentiality, and Privacy, With Revised Points of Focus (2022).</em> <a href="https://www.aicpa-cima.com/resources/download/2017-trust-services-criteria-with-revised-points-of-focus-2022">https://www.aicpa-cima.com/resources/download/2017-trust-services-criteria-with-revised-points-of-focus-2022</a></p><p>[9] International Organization for Standardization. <em>ISO/IEC 27001:2022 — Information Security, Cybersecurity and Privacy Protection — Information Security Management Systems — Requirements.</em> <a href="https://www.iso.org/standard/27001">https://www.iso.org/standard/27001</a></p><p>[10] UpGuard. <em>ISO 27001 Annex A Control 8.2: Privileged Access Rights.</em> Secondary control explanation and mapping. <a href="https://www.upguard.com/compliance/iso-27001/8-2">https://www.upguard.com/compliance/iso-27001/8-2</a></p><p>[11] Electronic Code of Federal Regulations. <em>45 CFR §164.312 — Technical Safeguards.</em> <a href="https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.312">https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.312</a></p><p>[12] Electronic Code of Federal Regulations. <em>45 CFR §164.308 — Administrative Safeguards.</em> <a href="https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.308">https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-C/section-164.308</a></p><p>[13] U.S. Department of Health and Human Services. <em>Summary of the HIPAA Security Rule.</em> <a href="https://www.hhs.gov/hipaa/for-professionals/security/laws-regulations/index.html">https://www.hhs.gov/hipaa/for-professionals/security/laws-regulations/index.html</a></p><p>[14] U.S. Office of Information and Regulatory Affairs. <em>HIPAA Security Rule Update, RIN 0945-AA22.</em> Federal Regulatory Agenda. <a href="https://mobile.reginfo.gov/public/do/eAgendaViewRule?RIN=0945-AA22&amp;pubId=202510">https://mobile.reginfo.gov/public/do/eAgendaViewRule?RIN=0945-AA22&amp;pubId=202510</a></p><p>[15] Apple Developer Documentation. <em>Configuring Multiple Devices Using Profiles.</em> <a href="https://developer.apple.com/documentation/devicemanagement/configuring-multiple-devices-using-profiles">https://developer.apple.com/documentation/devicemanagement/configuring-multiple-devices-using-profiles</a></p><p>[16] AuditFront. <em>SOC 2 Common Criteria CC6.3.</em> Secondary interpretation. <a href="https://www.auditfront.com/frameworks/soc-2/common-criteria/cc6-3/">https://www.auditfront.com/frameworks/soc-2/common-criteria/cc6-3/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd6d78d0a8d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Your Mac Update Reminder Needs a Second Voice]]></title>
            <link>https://medium.com/@raell.dottin/when-your-mac-update-reminder-needs-a-second-voice-3eb4a6727ecf?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/3eb4a6727ecf</guid>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-11-11T17:01:02.694Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Using DDM OS Reminder across a managed Mac fleet — and deciding whether its extra prompts improve the update experience.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NM30XpJ20fSaS-1f" /><figcaption>Photo by <a href="https://unsplash.com/@hullabaloo22?utm_source=medium&amp;utm_medium=referral">dianne clifford</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>An employee has a video call at four in the afternoon. Their Mac has a mandatory operating system update scheduled for six, and they still have documents open that they intend to finish before leaving work. The administrator who scheduled the update can see that the management policy was delivered, but that doesn’t establish whether the employee understands what the deadline will mean for their work.</p><p>At enforcement, macOS can force-quit running applications, including those containing unsaved documents, and restart the computer to install the required update. Apple documents this behavior as part of its managed software-update process. [9]</p><p>The operating system also provides notifications before enforcement. They become more persistent as the deadline approaches, and Apple allows organizations to provide additional information about the update. During the final 24 hours, its deadline notifications can bypass Do Not Disturb. [9]</p><p>That existing functionality is worth acknowledging because it changes how we should evaluate a separate reminder utility. Apple already has a mechanism for communicating mandatory updates. An organization considering another one needs a reason beyond wanting a larger notification window.</p><p>Dan Snelson’s <a href="https://github.com/dan-snelson/DDM-OS-Reminder">DDM OS Reminder</a> provides customizable employee messaging around operating system deadlines established through Apple’s Declarative Device Management. It can display the required version, remaining time, instructions, support information, and additional warnings while adjusting its behavior as the deadline approaches. [1]</p><p>The utility is an interesting solution to a legitimate operational problem, but deploying it should be a conditional decision. <strong>An additional reminder system is worthwhile when the existing notification experience leaves a demonstrated communication gap and the new system improves employee preparedness without introducing disproportionate interruptions, administrative complexity, or misleading information.</strong></p><p>That gives us something more useful to evaluate than whether the reminder window looks better than Apple’s notification.</p><h3>Apple Owns the Update Deadline</h3><p>Organizations generally manage Mac computers through a mobile device management service, or MDM. The term includes computers despite the word <em>mobile</em>. Products such as Jamf Pro and Microsoft Intune allow administrators to configure managed devices, distribute software, establish security requirements, and control operating system updates. [11] [12]</p><p>Apple’s Declarative Device Management, or DDM, allows a management service to describe a desired device state and leave more of the work of reaching that state to the device itself. For software updates, the management service can specify which operating system version is required and the deadline for installing it. The Mac receives that declaration and handles preparation, installation, and enforcement according to Apple’s rules. [3]</p><p>A software-update enforcement declaration contains fields such as TargetOSVersion, identifying the required version, and TargetLocalDateTime, specifying when enforcement should occur. The deadline is expressed in the device&#39;s local time without a time-zone offset. Consequently, an enforcement policy specifying six in the evening can apply at six o&#39;clock local time to Macs in different geographic regions. [4]</p><p>Once the requirement is active, macOS begins preparing the update and informing the employee. The employee can install it before the deadline instead of waiting for enforcement. If the deadline arrives without a completed installation, macOS attempts to perform the required update. [9]</p><p>There are circumstances in which that attempt cannot finish. The computer might be offline, lack sufficient storage, or fail another installation requirement. On an Apple silicon Mac, authorization may involve a bootstrap token, a credential maintained through device management that permits certain privileged operations without requiring the employee to enter their password. If that authorization isn’t available, the user may need to authenticate. [14]</p><p>Apple’s declarative system can continue pursuing the required operating system state when conditions permit. A missed enforcement time doesn’t necessarily mean the device has forgotten its obligation to update. [9]</p><p>This establishes the authority behind the process. The device-management service supplies the requirement, and macOS is responsible for acting on it. DDM OS Reminder doesn’t take ownership of that relationship.</p><h3>What DDM OS Reminder Adds</h3><p>DDM OS Reminder runs locally on the Mac and examines information that macOS records about managed software updates.</p><p>One source is /var/log/install.log, the system log containing software installation and update activity. The utility examines relevant entries to determine whether the Mac has a pending operating system requirement and when that requirement should be enforced. [1]</p><p>Log files contain historical information, however, and not every recorded declaration necessarily represents the policy currently governing the device.</p><p>Starting with version 5.0.0, the project added a declaration-aware resolution process. When Apple’s protected software-update declaration state is available and recognizable, the utility checks whether relevant installation-log entries correspond to an active declaration recorded by the operating system. [1]</p><p>This safeguard is more important than it might initially appear. A reminder utility that trusts an obsolete deadline could tell an employee that their computer is about to restart when the management service has already changed the policy.</p><p>Once the utility resolves a valid requirement, it compares the installed macOS version with the required version and decides whether communication is necessary. A Mac that already satisfies the requirement should not continue displaying update warnings.</p><p>For a computer that still needs the update, the interface is provided by <a href="https://github.com/swiftDialog/swiftDialog">swiftDialog</a>, an open-source macOS utility for presenting customizable windows from scripts. Administrators can supply their organization’s branding, instructions, support contacts, and actions such as opening Software Update. [1] [10]</p><p>The reminder is scheduled through a LaunchDaemon, a macOS background service that can operate independently of an employee opening an application. The project’s lightweight starter checks approximately once per minute whether another reminder is due and launches the main script when the scheduler requires it. A one-minute check interval does not mean the employee receives a message every minute. [5]</p><p>The architecture gives administrators control over the communication experience while leaving software installation under Apple’s control. It also introduces a second system whose interpretation of the update requirement must remain consistent with what macOS actually knows.</p><p>That consistency is where much of the operational work lies.</p><h3>Decide Whether Apple’s Notifications Are Enough</h3><p>Apple’s notification behavior is not limited to an occasional banner that employees can dismiss indefinitely. Managed update notifications become more frequent as enforcement approaches, and Apple provides an option on supported operating systems to reduce notifications so they appear only during the final hour, together with the restart countdown. [9]</p><p>An organization may find the native experience entirely adequate. If employees generally install updates before enforcement, understand the existing notifications, and experience few unexpected interruptions, another agent could add more complexity than value.</p><p>The case for DDM OS Reminder becomes stronger when an organization can identify something the existing experience doesn’t provide effectively.</p><p>That might be organization-specific instructions explaining how to prepare for an update, a direct link to internal support documentation, a localized message appropriate for a particular workforce, or a reminder schedule coordinated with common working hours. The utility supports these kinds of customizations through managed configuration settings. [1] [6]</p><p>Those capabilities are real, but their benefits are not automatic. A custom window can still be dismissed without being understood, and a frequent reminder can become another interruption that employees learn to ignore.</p><p>The decision should begin with evidence from the existing rollout. An administrator could review the percentage of devices completing updates before enforcement, the number of update-related support requests, and reports of employees being interrupted unexpectedly. These observations would help establish whether communication is a plausible source of the problem.</p><p>They would also help distinguish communication failures from installation failures. An employee who understands the deadline but cannot install the update because the Mac lacks storage is experiencing a different problem from one who never understood that an update was mandatory.</p><p>DDM OS Reminder can warn about low storage, but it cannot provide the missing space. The underlying operational problem still needs to be resolved.</p><h3>Build a Deployment You Can Verify</h3><p>The repository includes an assembly script that generates organization-specific deployment files. It is designed to work with management services capable of delivering configuration settings and executing scripts with administrative privileges. [5]</p><p>For this walkthrough, I’ll use com.example as a fictional organization&#39;s reverse-domain identifier. This identifier keeps the reminder&#39;s settings, background service, and files consistently named.</p><p>A production deployment should begin with a test Mac that has an actual pending DDM-enforced update. The Mac should also have a supported operating system, an interactive user session, and access to the resources required by the update and reminder.</p><p>Version 5.2.0 requires swiftDialog 3.1.0.4994 or later. The current swiftDialog 3.x release family supports macOS 15 and later, so organizations managing older Macs must review compatibility before deploying this particular reminder stack. Apple&#39;s declarative software-update capabilities can cover earlier supported macOS versions, but that doesn&#39;t make every third-party component compatible with them. [5] [10]</p><h3>Generate the test artifacts</h3><p>Start by cloning the repository and selecting the release used for this article:</p><pre>git clone https://github.com/dan-snelson/DDM-OS-Reminder.git<br>cd DDM-OS-Reminder<br><br>git checkout v5.2.0</pre><p>Pinning a release avoids silently incorporating changes made to the project’s development branch after the procedure has been reviewed.</p><p>The assembly script documents its supported options:</p><pre>zsh assemble.zsh --help</pre><p>Generate an interactive test configuration:</p><pre>zsh assemble.zsh com.example --lane test --interactive</pre><p>The script asks for organization-specific details such as support contacts, branding, optional restart behavior, and other reminder settings. It produces three deployment files in the Artifacts directory: an assembled installation script, a .plist property-list file containing preferences, and a .mobileconfig configuration profile.</p><p>A property list is Apple’s structured settings-file format. A configuration profile packages settings for installation through device management.</p><p>The project generates both formats to support different management workflows. Administrators should select one preference-delivery method rather than install both sets of settings. [5]</p><p>Before deploying the generated files, validate their syntax. On a clean build containing the newly generated test artifacts, the checks can be performed with:</p><pre>/usr/bin/plutil -lint Artifacts/com.example.dorm-*-test.plist<br><br>/usr/bin/plutil -lint \<br>  Artifacts/com.example.dorm-*-test-unsigned.mobileconfig</pre><p>plutil is Apple&#39;s command-line utility for working with property lists. Its lint operation checks the file&#39;s structure, although successful validation doesn&#39;t establish that every preference value is appropriate for the organization.</p><p>The administrator should also review the generated settings, particularly support links, reminder timing, restart behavior, and localization. A configuration can be syntactically valid and still be operationally unsuitable.</p><h3>Deploy the preferences before the script</h3><p>Upload the selected configuration to the management service. When using managed preferences, the domain must match the generated configuration, in this case com.example.dorm.</p><p>Once those preferences have reached the test Mac, execute the generated installation script through the management service with administrative privileges.</p><p>The installer creates or updates the reminder scripts and the LaunchDaemon responsible for scheduling their execution. The project validates generated runtime scripts before replacing installed copies, reducing the risk of leaving a partially written runtime behind during a deployment. [5]</p><p>The installed background service can be inspected using:</p><pre>sudo /bin/launchctl print system/com.example.dor</pre><p>Then verify the runtime files:</p><pre>ls -l /Library/Management/com.example/dor.zsh<br>ls -l /Library/Management/com.example/dor-starter.zsh</pre><p>The project’s operational log, using the standard path for this example, can be examined with:</p><pre>sudo tail -100 /var/log/com.example.log</pre><p>These checks show whether the expected components exist and what the utility reports. They don’t prove that the user-facing workflow behaves correctly.</p><p>The repository provides a demonstration mode for inspecting the interface:</p><pre>zsh reminderDialog.zsh demo</pre><p>Demonstration mode deliberately bypasses normal DDM state checks. It is useful for reviewing text, layout, and branding, but it cannot prove that the computer has a valid update requirement or that the production scheduler will display a reminder at the correct time. [5]</p><p>A separate runtime test on an approved Mac with a real pending declaration can evaluate the installed script:</p><pre>sudo /Library/Management/com.example/dor.zsh</pre><p>According to the project documentation, this manual invocation exercises production runtime logic without taking ownership of the normal daemon scheduler state. Scheduled execution must still be tested separately. [5]</p><p>This separation prevents a successful visual demonstration from being mistaken for a successful fleet deployment.</p><h3>Choose a Reminder Schedule Deliberately</h3><p>One detail worth catching before production deployment is the difference between the project’s built-in defaults and its generated sample configuration.</p><p>The configuration reference lists built-in thresholds of 60 days before enforcement for standard reminders, 45 days for more prominent presentation, and 21 days for restricting the secondary button. [6]</p><p>The current bundled Resources/sample.plist, however, explicitly sets those thresholds to 14, 3, and 1 day, respectively. [15]</p><p>Those are different configuration layers. The values in an actual deployed preference file take precedence over the corresponding script defaults, so an administrator should inspect the generated artifacts rather than assume that every reference to a default describes the behavior employees will see.</p><p>The sample also specifies baseline reminder times of 8 a.m., noon, and 4 p.m. in local time. Separate final-minute thresholds are configured at 45, 30, 15, 10, and 5 minutes before the effective deadline. [1] [15]</p><p>The frequency and visibility of these reminders should reflect the organization’s policy and the disruption associated with the update.</p><p>For example, an employee-facing Mac that regularly runs presentation software may benefit from early instructions explaining how to prepare for a restart. An unattended workstation may have little reason to display that same kind of interactive message. The two devices can require the same operating system version while needing different communication treatment.</p><p>DDM OS Reminder includes a mechanism for delaying notifications when specified conferencing applications appear to be active. It uses system activity information, including display-sleep assertions associated with applications such as Microsoft Teams, Zoom, and Webex, rather than reading an employee’s actual meeting schedule. [6] [15]</p><p>That implementation is useful, but the distinction affects how it should be tested. A running conferencing application doesn’t necessarily indicate an active meeting, and the absence of a recognized assertion doesn’t prove that an employee is available to be interrupted.</p><p>The utility also recognizes certain confirmed <strong>Update Tonight</strong> schedules and can temporarily suppress ordinary reminders. Final-minute thresholds and other urgent behaviors may still apply. [1]</p><p>Once the configured enforcement deadline has passed, the project’s aggressive reminder mode is enabled by default. Its standard settings begin urgent redisplay two hours after the deadline and repeat it every 20 minutes while the Mac remains below the required version. [6]</p><p>Administrators can adjust those settings and temporarily suppress aggressive mode using the project’s support mechanism. This should be tested against actual employee workflows rather than treated as an inherently better policy.</p><p>The optional post-deadline restart behavior deserves particular attention. DDM OS Reminder supports Off, Prompt, and Force settings, with Off as the default. These controls belong to the reminder utility&#39;s separate restart workflow; they are not replacements for Apple&#39;s update-enforcement policy. [1]</p><p>I would keep that behavior disabled during an initial pilot unless there is a documented requirement for it. Apple’s existing enforcement is already capable of interrupting employees, and adding another restart mechanism increases the number of interactions an administrator must understand and troubleshoot.</p><h3>When the Reminder and Apple’s Policy Disagree</h3><p>The utility’s behavior when a valid requirement cannot be established is among its most important design decisions.</p><p>A missing reminder may indicate that the Mac is already updated, that the next reminder isn’t due, or that an employee interaction has triggered a quiet period. It may also mean that the script could not resolve a trustworthy DDM declaration.</p><p>The resolver distinguishes conditions including missing, conflict, noMatch, and invalidVersion. Under normal operation, unresolved requirements result in suppression rather than the creation of an unsupported deadline. [1]</p><p>Version 5.2.0 makes those outcomes easier to investigate. Its client-side logging now uses entries such as DDM update check:, No reminder shown:, and Next step: to explain what happened and what an administrator should inspect next. [2]</p><p>The project also includes a read-only diagnostic helper, diagnoseReminderState.zsh, which can examine collected client logs, installation logs, and preference files to construct a more complete explanation of reminder behavior. [7]</p><p>There is an optional complication: DDM Emergency Fallback.</p><p>The project allows administrators to provide a required macOS version and deadline through deployment parameters when normal declaration resolution is unavailable or unreliable. A confirmed Apple declaration takes precedence, and invalid fallback information is rejected. The fallback is considered for documented unresolved states rather than replacing normal resolution unconditionally. [7]</p><p>Imagine that a Mac’s local update information conflicts with the active declarations expected by the administrator. Normal resolution cannot establish a trustworthy requirement, so the reminder remains silent.</p><p>The administrator could deploy an emergency fallback specifying a version and deadline. If the fallback is valid, the reminder may resume displaying messages based on that information even though the normal Apple declaration was not confirmed.</p><p>That creates a communication risk.</p><p>The employee might see a deadline that appears to carry the same authority as an Apple-enforced update, although the reminder is using a separately supplied requirement. The actual operating system could have a different enforcement state, and the fallback itself does not create an Apple declaration.</p><p>This is why I would require an administrator to verify the intended version and deadline against the management service before activating fallback messaging. If Apple’s enforcement status remains uncertain, employee-facing language should not confidently promise that macOS will perform an automatic restart at the fallback time. That is an operational safeguard I recommend, not a claim that the tool automatically rewrites its messages in this way.</p><p>Fallback should have an identified owner and a plan for removing or replacing it once the original declaration problem is resolved. Otherwise, a temporary workaround risks becoming another source of policy information that administrators must maintain.</p><p>The security policy identifies a related limitation. When Apple’s protected declaration state is unavailable or unrecognized, the project can continue using installation-log evidence for compatibility. Because local users can append to that log, uncorroborated entries can potentially produce incorrect reminders or inventory information. They cannot modify Apple’s actual DDM enforcement through the reminder utility. [8]</p><p>The same principle applies to the project’s low-battery and low-storage warnings. Version 5.2.0 adds a battery warning when a Mac is using its internal battery below the configured threshold, which defaults to 50 percent. The generated sample’s disk-space warning threshold is 35 gigabytes. Both are communication features, not guarantees that the update can install successfully. [2]</p><p>These behaviors illustrate why message accuracy is an operational responsibility. The reminder’s content needs to be supported by trustworthy device state, and the language should not claim more authority than the evidence provides.</p><h3>Test Whether the Additional Reminders Improve Anything</h3><p>A pilot should establish more than whether DDM OS Reminder launches and displays its interface.</p><p>I would begin by selecting comparable groups of managed Macs that are subject to the same required operating system version and enforcement policy. One group would continue using Apple’s native notifications; the other would receive the additional reminder utility.</p><p>The comparison needs to account for differences between the groups. A population of office-based engineers on recent Apple silicon laptops may behave differently from a population of remote employees using older computers with limited storage. If those conditions vary substantially, a difference in completion rates cannot confidently be attributed to the reminders.</p><p>The most useful primary outcome is the share of eligible Macs that reach the required operating system version before enforcement. That requires verified operating system inventory, ideally including the build number where the update policy requires a particular build.</p><p>Jamf Pro provides update-management capabilities, while Microsoft Intune exposes Apple software-update reporting that includes current and pending operating system information, installation states, and failures. Administrators should use the appropriate management platform’s documented reporting rather than infer compliance from reminder logs. [11] [12] [16]</p><p>A second outcome is employee disruption. Help-desk tickets about unexpected restarts, employee reports of poorly timed interruptions, and complaints about repeated reminders can indicate whether the communication experience has improved or deteriorated.</p><p>Those measures are imperfect. Ticket volume depends on how people report problems, and employee surveys depend on response rates and expectations. They become more useful when the organization applies comparable measurement methods before and during the pilot.</p><p>I would also track the reliability of the additional component itself. Failed script deployments, missing reminder schedules, inaccurate version information, unexpected dialogs, and recurring support interventions are costs of operating the system.</p><p>The evaluation should answer three connected questions. Did the reminder reach the right employee with accurate information? Did that information help the employee complete the update before enforcement? Was the improvement sufficient to justify maintaining another privileged background component across the fleet?</p><p>A positive result on the first question does not establish the second. Employees may notice an additional dialog without changing when they install the update. Likewise, a higher update-completion rate during the pilot would be encouraging but would require careful interpretation if the groups differed in hardware, network availability, or management policy.</p><p>Before beginning the pilot, the organization should specify what improvement would justify adoption and what level of additional interruption or support burden it considers unacceptable. Without those criteria, nearly any outcome can be described as success after the fact.</p><p>For some organizations, the result may be that Apple’s native notifications are already sufficient. That would be a useful finding because it avoids deploying and maintaining an unnecessary component.</p><h3>Maintaining the Reminder Is Part of the Decision</h3><p>An open-source utility installed across a Mac fleet introduces ongoing responsibilities beyond its initial configuration.</p><p>DDM OS Reminder runs with elevated privileges, maintains local state, reads software-update records, and uses swiftDialog to present its interface. The project’s security guidance recommends reviewing trusted distribution paths, configuration files, generated scripts, and release changes before production deployment. [8]</p><p>The update to version 5.2.0 provides a practical example. It introduced battery-warning placeholders in message templates and changed several diagnostic log messages. Administrators using customized message text must update the relevant templates, and scripts that parse older log wording may require adjustment. [2]</p><p>These changes are understandable, but they demonstrate why the project is not entirely maintenance-free.</p><p>A release may improve the employee experience while requiring coordinated updates to runtime scripts and managed preferences. A customized installation can also behave differently from the maintainer’s sample configuration, particularly when older settings are imported.</p><p>The latest assembly workflow can import supported previous .plist files and, beginning with version 5.2.0, appropriate .mobileconfig profiles. That helps preserve organizational customizations, but the resulting artifacts still need review. [2] [7]</p><p>I would manage the generated configuration and deployment script as a versioned change, with test and production scopes, recorded release identifiers, and a rollback procedure. The background service should be verified after deployment, and its diagnostic logs should be available when administrators investigate unexpected behavior.</p><p>Organizations should also consider how they will retire the utility if they no longer need it. The project’s installer supports controlled uninstallation of the runtime, while removal of separately managed preferences remains an administrative responsibility. [5]</p><p>Those considerations are part of evaluating the cost of the additional communication layer. A tool that reduces employee interruptions but regularly creates deployment incidents may not represent a net improvement.</p><h3>The Decision I Would Make</h3><p>DDM OS Reminder addresses an area of Mac management that deserves more attention: employees need reliable information about maintenance that can interrupt their work, even when the underlying enforcement policy is technically correct.</p><p>The utility supplies useful capabilities for organizations that need customized reminders, support instructions, localization, and more control over notification timing. Its approach of interpreting locally available DDM state, rather than taking over software installation, provides a practical way to build on Apple’s existing management architecture. [1]</p><p>I would not make it a standard component of every managed Mac solely because those features exist.</p><p>The justification should come from the rollout being managed. If native notifications already lead to timely updates with few employee complaints, there may be no communication problem for the utility to solve. If employees frequently miss the preparation window, misunderstand the deadline, or need support instructions that the native experience does not communicate effectively, a controlled pilot is reasonable.</p><p>The pilot should preserve Apple’s enforcement policy, keep reminder-derived information separate from confirmed enforcement state, and evaluate both update completion and employee disruption. The administrator should also account for the security, compatibility, and maintenance costs of the additional software.</p><p>The broader lesson extends beyond this particular project. A managed update has at least three states worth keeping separate: the version and deadline the organization requires, the information the employee receives, and the operating system state the computer eventually reaches. Each can succeed or fail independently.</p><p>A correct policy doesn’t establish that the employee was prepared. A reminder displayed on screen doesn’t establish that the update was installed. And an installed update doesn’t tell us whether the organization could have completed the process with less disruption.</p><p>Keeping those observations separate gives administrators a better basis for deciding what to improve. DDM OS Reminder is worth considering where that evaluation identifies a communication problem it can actually address.</p><h3>References</h3><p>[1] Dan K. Snelson. <em>DDM OS Reminder — README</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/README.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/README.md</a></p><p>[2] Dan K. Snelson. <em>DDM OS Reminder, Version 5.2.0 Release Notes</em>. October 20, 2026. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/releases/tag/v5.2.0">https://github.com/dan-snelson/DDM-OS-Reminder/releases/tag/v5.2.0</a></p><p>[3] Apple Developer Documentation. <em>Deploying Software Updates Using Declarative Management</em>. <a href="https://developer.apple.com/documentation/devicemanagement/deploying-software-updates-using-declarative-management">https://developer.apple.com/documentation/devicemanagement/deploying-software-updates-using-declarative-management</a></p><p>[4] Apple Developer Documentation. <em>SoftwareUpdateEnforcementSpecific</em>. <a href="https://developer.apple.com/documentation/devicemanagement/softwareupdateenforcementspecific">https://developer.apple.com/documentation/devicemanagement/softwareupdateenforcementspecific</a></p><p>[5] DDM OS Reminder. <em>Deployment Workflow</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/04-deployment-workflow.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/04-deployment-workflow.md</a></p><p>[6] DDM OS Reminder. <em>Configuration Reference</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/06-configuration-reference.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/06-configuration-reference.md</a></p><p>[7] DDM OS Reminder. <em>Resources Documentation</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Resources/README.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Resources/README.md</a></p><p>[8] DDM OS Reminder. <em>Security Policy</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/SECURITY.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/SECURITY.md</a></p><p>[9] Apple. <em>Install and Enforce Software Updates for Apple Devices</em>. Apple Platform Deployment. <a href="https://support.apple.com/guide/deployment/install-and-enforce-software-updates-depd30715cbb/web">https://support.apple.com/guide/deployment/install-and-enforce-software-updates-depd30715cbb/web</a></p><p>[10] swiftDialog Project. <em>swiftDialog — Create User Notifications on macOS</em>. GitHub. <a href="https://github.com/swiftDialog/swiftDialog">https://github.com/swiftDialog/swiftDialog</a></p><p>[11] Jamf. <em>Updating macOS Using Managed Software Updates</em>. Jamf Pro Documentation. <a href="https://learn.jamf.com/r/en-US/jamf-pro-documentation-current/Updating_macOS_Using_Managed_Software_Updates">https://learn.jamf.com/r/en-US/jamf-pro-documentation-current/Updating_macOS_Using_Managed_Software_Updates</a></p><p>[12] Microsoft. <em>Configure Update Policies for Apple Devices</em>. Microsoft Intune Documentation. <a href="https://learn.microsoft.com/en-us/intune/device-updates/apple/">https://learn.microsoft.com/en-us/intune/device-updates/apple/</a></p><p>[13] DDM OS Reminder. <em>Configuration Hierarchy</em>. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/05-configuration-hierarchy.md">https://github.com/dan-snelson/DDM-OS-Reminder/blob/main/Diagrams/05-configuration-hierarchy.md</a></p><p>[14] Apple. <em>Use Secure Token, Bootstrap Token, and Volume Ownership in Deployments</em>. Apple Platform Deployment. <a href="https://support.apple.com/guide/deployment/use-secure-and-bootstrap-tokens-dep24dbdcf9e/web">https://support.apple.com/guide/deployment/use-secure-and-bootstrap-tokens-dep24dbdcf9e/web</a></p><p>[15] DDM OS Reminder. <em>Resources/sample.plist</em>, Version 5.2.0. GitHub. <a href="https://github.com/dan-snelson/DDM-OS-Reminder/blob/v5.2.0/Resources/sample.plist">https://github.com/dan-snelson/DDM-OS-Reminder/blob/v5.2.0/Resources/sample.plist</a></p><p>[16] Microsoft. <em>Software Update Reporting for Apple Devices</em>. Microsoft Intune Documentation. <a href="https://learn.microsoft.com/en-us/intune/device-updates/apple/monitor-reports">https://learn.microsoft.com/en-us/intune/device-updates/apple/monitor-reports</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3eb4a6727ecf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Wanted to Build the Next Gmail. Now I Want to Build Two Companies.]]></title>
            <link>https://medium.com/@raell.dottin/i-wanted-to-build-the-next-gmail-now-i-want-to-build-two-companies-989184321cb9?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/989184321cb9</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 17:01:02 GMT</pubDate>
            <atom:updated>2026-11-10T17:01:02.654Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>From an invitation-only email service to the practical questions of starting a software company and a quantitative trading firm.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XsLCfIjBlVqxgchW" /><figcaption>Photo by <a href="https://unsplash.com/@mdishakrahman?utm_source=medium&amp;utm_medium=referral">Md Ishak Rahman</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The first Gmail account I had came from an invitation a friend gave me. This was back when Gmail was still in beta, and getting an account depended on knowing someone who could invite you.</p><p>Google had introduced Gmail in April 2004, offering a gigabyte of storage at a time when other email providers were considerably less generous. The application also made searching, organizing, and navigating email more convenient than I was accustomed to. [1]</p><p>I was impressed by what Google had built, but I was also thinking about what I might build myself.</p><p>One idea was an email service similar to Gmail with GNU Privacy Guard built directly into it. GNU Privacy Guard, usually called GnuPG or GPG, is free software that supports encryption and digital signatures. Encryption protects information by making it unreadable without the appropriate key. Digital signatures allow someone to verify that a message was signed by a particular key and hasn’t been altered. [2]</p><p>I imagined an email application where people could protect their messages without installing separate software or learning how to operate cryptographic tools.</p><p>I didn’t appreciate all the difficulties involved. A study published in 1999, <em>Why Johnny Can’t Encrypt</em>, examined whether people unfamiliar with cryptography could use an email encryption application. Most participants couldn’t successfully sign and encrypt a message within the allotted 90 minutes. [3]</p><p>That problem would have been central to my idea. An encrypted email service needs a reliable way to distribute public keys, verify recipients, and help users recover from mistakes. Encrypting a message’s contents also doesn’t necessarily conceal information about who sent it, who received it, or when it was delivered.</p><p>Making encryption available and making it usable are separate engineering responsibilities.</p><p>Around that time, I had another ambition: becoming the Bill Gates of mobile phones by developing an operating system popular enough to compete with the industry’s established companies.</p><p>An operating system is the software responsible for managing a device’s hardware and providing the foundation on which applications run. In the mobile market, creating one would have involved considerably more than writing the system software. Device manufacturers, application developers, wireless carriers, distribution, updates, and hardware compatibility would all have influenced whether people could actually use it.</p><p>I was interested in creating technology that people would depend on, although I knew relatively little about establishing the organizations needed to develop and maintain it.</p><p>Years later, I’m approaching the problem from a different position. I have mobile application projects, including Owlory, a planning application, and Gymphant, a workout application. I’m interested in developing additional mobile applications and software-as-a-service products. Software as a service, or SaaS, generally means providing software as an ongoing service, often over the internet and through recurring subscriptions.</p><p>I’m also developing my interest in quantitative trading, where mathematical models and software help determine when financial instruments should be bought or sold.</p><p>I want a company for the software products and, eventually, a separate proprietary trading firm for the quantitative systems.</p><p>The two businesses don’t need to be established on the same timetable. My applications already present questions about ownership, publishing, customer relationships, and operating expenses. The trading project has a more fundamental question to answer first: whether its models can be tested and operated reliably enough to justify committing company capital.</p><p>I’ve come to see company formation as a decision that should follow the work a business needs to perform. A legal entity becomes useful when it provides a clearer way to own assets, enter agreements, manage obligations, or contain particular risks. Creating the entity before those requirements exist can introduce expenses without solving much.</p><p>I wanted to understand how other technology companies reached those decisions before making my own.</p><h3>How the Companies I Grew Up Admiring Actually Started</h3><p>Apple is an interesting example because the company existed before it incorporated.</p><p>Steve Jobs, Steve Wozniak, and Ronald Wayne established Apple Computer Company as a partnership in 1976. Wayne left shortly afterward, and Mike Markkula later became involved as an investor and business adviser. Apple incorporated on January 3, 1977. [4]</p><p>A general partnership allows people to conduct a business together without first establishing a corporation. The arrangement can be comparatively simple, although partners may face personal responsibility for business obligations.</p><p>Apple’s move toward incorporation occurred as the business was developing beyond its original partnership and preparing for greater manufacturing, financing, and commercial activity. That sequence is relevant to my research, but it doesn’t establish which particular requirement made incorporation necessary or whether another structure could have worked.</p><p>Microsoft also began before its corporate formation.</p><p>Bill Gates and Paul Allen started working together commercially in 1975, developing software for the Altair microcomputer. Microsoft’s historical records describe their early partnership and software licensing activities. Microsoft became a Washington corporation on June 25, 1981. It subsequently reincorporated in Delaware in 1986 and returned to Washington in 1993. [5][6]</p><p>There were six years between Microsoft’s first commercial activities and incorporation. That doesn’t mean waiting six years is advisable. It shows that a company can develop products, establish customer relationships, and generate revenue before adopting the corporate structure it will eventually use.</p><p>Google and Amazon followed different paths.</p><p>Google incorporated in California in September 1998 and reincorporated in Delaware in August 2003, before its 2004 initial public offering. Amazon incorporated in Washington in 1994, moved its incorporation to Delaware in 1996, and became publicly traded in 1997. [7][8]</p><p>Both began as corporations rather than first establishing limited liability companies.</p><p>An initial public offering, commonly abbreviated IPO, occurs when a company first offers its stock to public investors. Preparing for one creates requirements involving ownership, disclosures, governance, and securities regulation that an independent application developer may never encounter.</p><p>The company history that most directly answers my original question is Workiva.</p><h3>What Workiva’s Conversion Actually Tells Us</h3><p>Workiva began in California in August 2008 as WebFilings LLC. It developed cloud-based software that helped businesses collect, manage, and report financial and operational information.</p><p>In 2014, the company became Workiva LLC, converted from a California limited liability company to a Delaware limited liability company, and then converted into Workiva Inc., a Delaware corporation, on December 10, shortly before its public offering. [9]</p><p>The conversion involved considerably more than filing paperwork under a different business name.</p><p>Workiva had several classes of ownership interests, including preferred units, common units, and options. Its corporate conversion replaced those interests with shares of stock or options to purchase shares according to documented conversion formulas. [9]</p><p>A <em>limited liability company</em>, or LLC, represents ownership through membership interests. A corporation ordinarily represents ownership through stock. Converting between the two can require adjustments to investors’ rights, compensation arrangements, tax treatment, and ownership records.</p><p>Workiva’s disclosures also explain that it was taxed as a pass-through entity before its corporate conversion and became subject to corporate income taxation afterward. [9]</p><p>That is a useful example because the legal change accompanied a specific financial event. The company was preparing to sell shares to public investors, and its existing ownership interests needed to fit the corporate structure being used for that offering.</p><p>It doesn’t follow that every LLC should convert once it grows beyond a certain revenue level. Workiva’s financing and ownership requirements are evidence for understanding its decision, not instructions for my own company.</p><p>Large businesses can also continue operating through LLC entities. Jane Street, a major quantitative trading organization, uses limited liability company entities within its business structure. Its regulatory filings describe sophisticated trading and financial activities, although they don’t provide a ready-made organizational plan for an independent trader. [10]</p><p>These examples changed how I think about choosing an entity. I was initially interested in what successful founders registered. The more useful questions concern what their businesses were doing, who needed to own them, and what their structures had to accommodate.</p><h3>What Would My Software Company Need to Own?</h3><p>For the software business, the first decision is whether formalizing the organization would solve problems that exist now.</p><p>An individual developer can create and sell applications without forming a separate LLC or corporation. Operating as a sole proprietor is a legitimate possibility, particularly when the software is experimental or generates little revenue. The U.S. Small Business Administration recognizes different business structures because their usefulness depends on ownership, financing, taxation, and liability requirements. [11]</p><p>I don’t need a corporation merely to write Swift code, run an application on my iPhone, or experiment with a subscription feature.</p><p>My existing projects raise more concrete questions, however.</p><p>Owlory and Gymphant have code, names, developer accounts, and ongoing maintenance requirements. As I consider publishing and developing them through a company, I need to establish which assets the company would own and which obligations it would assume.</p><p>Creating an LLC wouldn’t automatically transfer software I previously developed as an individual into the company. Ownership of existing code, trademarks, domain names, and related assets would need to be documented appropriately.</p><p>If another developer eventually contributes to Gymphant, for example, I would want a written agreement covering payment, ownership of contributed work, confidentiality, and responsibility for delivering it. If Owlory develops recurring revenue, the business needs accounting that separates application income and expenses from my personal finances.</p><p>The publishing relationship also requires attention. Apple’s App Store displays a seller name associated with the developer’s legal status. If I want the company to be the publisher, its legal identity and developer membership have to be addressed.</p><p>These are reasons to investigate formation now, even though I haven’t established that incorporation is immediately required for every application.</p><p>An LLC could provide a legal owner for the software, a counterparty for agreements, and a way to organize business finances. Its limited liability features may protect an owner’s personal assets from some company obligations, provided the entity is properly maintained and subject to important exceptions such as personal guarantees and misconduct.</p><p>I would still compare those advantages with remaining a sole proprietor for a period. Publication requirements, legal assistance, annual filings, tax preparation, and other administrative costs don’t disappear simply because an application has few customers.</p><p>The timing should depend on when the benefits become substantial enough to justify those obligations.</p><h3>LLC, Corporation, and the New York Question</h3><p>A single-owner LLC is ordinarily treated as part of its owner’s federal income-tax return unless it elects another classification. An LLC with multiple owners is generally taxed as a partnership by default. These arrangements are often called <em>pass-through taxation</em>, because income is generally reported by the owners rather than being subject to ordinary federal corporate income tax at a separate entity level. [12]</p><p>The state-law existence of an LLC and its federal income-tax classification are different questions. A single-member LLC may be disregarded as separate from its owner for federal income-tax purposes while remaining a separate legal entity under state law.</p><p>A corporation offers a familiar structure for issuing shares, appointing directors, and establishing ownership rights among founders and investors.</p><p>A C corporation is generally taxed separately from its shareholders. An S corporation is an eligible corporation, or other qualifying entity, that elects a special federal tax treatment under rules that restrict matters such as eligible shareholders and classes of stock. These terms describe federal tax treatment rather than additional forms of incorporation under New York law. [35]</p><p>For my software business, the practical comparison is likely to be a New York LLC and a corporation, with Delaware incorporation becoming relevant if financing or ownership requirements justify it.</p><p>A venture-backed company may benefit from adopting a corporate structure early, particularly when investors expect shares, preferred stock, or employee stock options. If I intend to retain ownership and fund development through customer revenue, those advantages may not justify the additional complexity.</p><p>New York’s filing costs are straightforward enough to verify.</p><p>The Department of State charges $200 to file the Articles of Organization for a domestic LLC. It charges $125 for the Certificate of Incorporation for a domestic business corporation. [13][14]</p><p>The LLC has an additional publication requirement. Most newly formed New York LLCs must publish a formation notice in two designated newspapers for six consecutive weeks. The company then files a Certificate of Publication, together with the publishers’ affidavits, for a $50 state filing fee. Missing the 120-day deadline can result in suspension of the LLC’s authority to conduct business. Newspaper advertising costs are separate and vary. [13]</p><p>Both types of entity generally must file a biennial statement every two years, currently costing $9. Depending on its classification and New York income, an LLC may also have an annual tax-related filing-fee obligation. [15][20]</p><p>Those figures establish the government filing expenses, not the total cost of operating the company.</p><p>I would also need to decide on an operating agreement or other governance documents, establish banking and bookkeeping, and obtain an Employer Identification Number where needed. An Employer Identification Number, or EIN, is the federal tax identifier issued free by the Internal Revenue Service. Even where a single-owner LLC is not required to obtain one for a particular federal income-tax purpose, an EIN may be useful or required for banking and other obligations. [16]</p><h3>Making the Company the App Store Seller</h3><p>Apple’s requirements make the abstract ownership question more practical.</p><p>To enroll an organization in the Apple Developer Program, Apple generally requires proof that the organization is a legal entity, evidence that the applicant has authority to enter agreements on its behalf, and a D-U-N-S Number. D-U-N-S refers to the Data Universal Numbering System, a business identification system maintained by Dun &amp; Bradstreet. [17]</p><p>Apple also supports converting an eligible individual developer membership into an organization membership through an account review process. Consequently, forming a company doesn’t necessarily require creating an unrelated developer account and transferring every application into it. [34]</p><p>If a transfer is necessary, Apple’s application-transfer process can preserve an application’s identity, ratings, and reviews, subject to eligibility and additional work for capabilities such as subscriptions and certain cloud services. [18]</p><p>I would investigate the membership-conversion route before changing anything that customers already depend on.</p><p>There is also the question of software sales tax.</p><p>New York generally treats the sale or licensing of prewritten software as taxable, including qualifying software accessed remotely over the internet. That can affect a SaaS product even though the customer never downloads a conventional installation package. The exact treatment depends on the product, transaction, and where the software is used. [19]</p><p>For me, that means the company needs to know which sales are made through Apple, which might eventually be made directly, and who is responsible for collecting and reporting applicable taxes.</p><p>The value of legal and accounting advice becomes clearer at this point. New York City’s Department of Small Business Services offers legal consultations to qualifying small businesses and entrepreneurs on formation, contracts, intellectual property, and related matters. [33]</p><p>I would prefer to resolve those questions before transferring application ownership or signing contracts in the company’s name.</p><h3>Why Older Formation Guides Need Checking</h3><p>The administrative requirements can change after an article or tutorial is published.</p><p>Delaware, for example, increased its annual LLC tax from $300 to $400 for the 2026 tax year under House Bill 400, with the higher tax first generally payable in 2027. A guide using the old amount could lead someone to underestimate continuing expenses. [22]</p><p>Beneficial ownership reporting has changed as well. Under federal rules announced in March 2025, companies created in the United States are exempt from the Financial Crimes Enforcement Network’s federal beneficial ownership reporting requirements. New York’s reporting framework, effective in 2026, applies to qualifying LLCs formed under the law of a foreign country and authorized to do business in New York. [23][24]</p><p>I wouldn’t choose an entity based on a checklist without confirming that the cited requirements still apply.</p><p>The formation decision needs to be grounded in the company’s activities and the law in effect when I register it.</p><h3>Quantitative Trading Has a Different Starting Point</h3><p>My proposed trading company requires another kind of analysis.</p><p>Quantitative trading uses mathematical and statistical methods to study financial markets and generate trading decisions. An automated system may take those decisions and submit orders through a brokerage interface.</p><p>The software might examine stock prices, trading volume, volatility, relationships among securities, or economic information. The model’s output could determine which instrument to trade, how much capital to allocate, and when to exit a position.</p><p>That is already an interesting engineering problem. Establishing a proprietary trading business introduces additional financial responsibilities.</p><p>A proprietary trading firm generally trades capital belonging to the firm and bears the resulting gains and losses. I am interested in building a business around trading its own capital rather than immediately raising money from outside investors.</p><p>Before treating that as an operating company, I need evidence that the trading software can function reliably under conditions resembling actual market activity.</p><p>A strategy may look attractive when tested against historical data. This process, called <em>backtesting</em>, applies a model to past market conditions to estimate what its results might have been.</p><p>The danger is that the simulation can appear more convincing than its evidence warrants.</p><p>A model might have been adjusted repeatedly until it performed well during the historical period used to develop it. That is a form of <em>overfitting</em>, where a model learns details of the available data that may not hold up in future conditions.</p><p>A backtest can also accidentally use information that wasn’t available at the time of a simulated decision. That problem is called <em>look-ahead bias</em>.</p><p>Execution introduces another source of error. The price visible when the model generates an order may differ from the price received when the order executes. This difference is called <em>slippage</em>. Commissions, financing expenses, bid-ask spreads, and market impact can further change the result.</p><p>The Securities and Exchange Commission’s investor guidance emphasizes that backtested results are hypothetical and should not be confused with actual historical investment performance. [32]</p><p>That warning applies even if the simulation was written carefully.</p><p>For my trading project, I would want to preserve the market data and model configuration used for each test, keep development data separate from later evaluation data, and compare simulated results with an appropriate benchmark after accounting for realistic expenses.</p><p>I would also want independent verification that the reported results can be reproduced. A model producing an attractive performance chart is not sufficient evidence that it can trade safely.</p><p>Paper trading would help test order generation and brokerage integration without risking actual funds. It is still a simulation, however, and may not reproduce real order fills, liquidity constraints, or market impact.</p><p>Live trading with a limited, explicitly risked amount of capital could eventually provide evidence that simulations cannot. That step would need to comply with brokerage requirements and applicable law, and it would require limits on exposure and losses.</p><p>The engineering controls are particularly important because automated systems can repeat mistakes quickly.</p><p>If a market-data feed stops updating, the trading system must avoid acting as though old prices are current. If an order is rejected, the program needs to understand that rejection before deciding whether to submit another. If the brokerage reports a different position from the one recorded internally, trading should not continue on the assumption that the internal record is correct.</p><p>The Financial Industry Regulatory Authority has published supervisory guidance on algorithmic trading systems, including development, testing, validation, and trading controls. Although that guidance addresses relevant regulated firms rather than prescribing identical requirements for every independent trader, the engineering failures it discusses are worth studying. [30]</p><p>I would want position limits, reconciliation against broker records, monitoring of orders and balances, and a way to stop new orders when the system enters an uncertain state.</p><p>There is a financial test alongside those engineering tests. The expected benefit of a strategy must justify its capital requirements, transaction expenses, operational costs, and potential losses.</p><p>A strategy that produces positive results in a backtest but cannot survive realistic costs doesn’t justify investment simply because the code runs correctly. Even a positive live track record wouldn’t guarantee its future profitability.</p><p>That is why the research and validation program should come before a substantial commitment to a proprietary trading operation.</p><h3>Trading My Own Money Doesn’t Settle Every Regulatory Question</h3><p>I also need to distinguish the type of trading business I want from other financial businesses that may appear similar.</p><p>A firm trading only its own capital is different from an investment adviser managing customer assets for compensation.</p><p>In securities markets, the Securities and Exchange Commission distinguishes traders from dealers. An ordinary trader buying and selling securities for their own account isn’t necessarily a registered dealer. A business that regularly performs dealer-like activities, such as certain forms of market making, may have different obligations. [25]</p><p>The analysis depends on what the business actually does rather than what its owner calls it.</p><p>If I eventually provide securities investment advice to others for compensation, investment adviser requirements could become relevant. [26]</p><p>Trading derivatives can introduce another regulatory framework. Providing advice about futures or other covered instruments may trigger commodity trading adviser requirements, while operating a pool of customer capital to trade covered instruments may trigger commodity pool operator rules. The National Futures Association provides registration guidance for those activities. [27][28]</p><p>Even proprietary securities activity can produce reporting obligations as trading volume increases. The SEC’s large-trader reporting framework is one example. [29]</p><p>I would therefore establish the intended instruments, capital arrangements, brokerage relationships, and trading activities before concluding what registration requirements apply.</p><p>Tax classification is a separate question. The Internal Revenue Service has criteria for determining when securities trading constitutes a trade or business for federal income-tax purposes, including whether trading is substantial, continuous, and regular. Qualifying traders may elect mark-to-market accounting under specific rules and deadlines. [31]</p><p><em>Mark-to-market accounting</em> generally recognizes gains and losses on covered positions based on their value at the end of the tax year, rather than waiting for an actual sale, subject to the applicable tax rules.</p><p>Forming an LLC doesn’t automatically qualify the owner for this treatment, and obtaining tax treatment as a trader doesn’t establish that the business is exempt from securities regulation.</p><p>This is where I would consult professionals familiar with proprietary trading, securities or derivatives regulation, and trader taxation before committing substantial live capital or representing the operation as a regulated financial firm.</p><h3>Should the Software and Trading Businesses Be Separate?</h3><p>My initial preference is to keep the two activities in separate companies once each has reached the point where a formal entity is justified.</p><p>The software company would own or license its applications, receive product revenue, maintain customer relationships, and pay expenses associated with software development.</p><p>The trading company would manage its authorized trading capital, maintain brokerage records, and assume the financial results of its strategies.</p><p>This separation would make it easier to evaluate the businesses independently.</p><p>Consider what happens if a software product produces subscription revenue while a trading strategy experiences a substantial drawdown. A <em>drawdown</em> is a decline in account value from a previous high point. I wouldn’t want funds reserved for software maintenance, customer obligations, and operating expenses casually exposed to losses from an unrelated trading model.</p><p>Separate companies can create clearer legal and accounting boundaries, provided their finances, contracts, and operations are actually kept separate. They do not eliminate the owner’s financial exposure from personal guarantees, misconduct, or other obligations.</p><p>There is also a cost to separation. Two entities may require separate tax returns or reporting, additional bookkeeping, professional services, and continuing administrative work. If the trading operation remains a research project, those expenses may be premature.</p><p>An alternative would be to establish the software business first and continue the trading research personally until its capital, operational, and regulatory requirements become clearer.</p><p>Another possibility is to operate related activities through a single entity initially, although that would need a careful review of liability, taxes, contracts, and risk exposure. I would be reluctant to expose the company’s software operations to trading risks without a compelling operational reason.</p><p>New York City’s Unincorporated Business Tax illustrates why the legal analysis can’t end with the organizational chart.</p><p>The city generally imposes a 4% tax on taxable income from covered unincorporated businesses. Qualifying nondealer activities involving buying, holding, and selling property for one’s own account can receive special treatment. The city’s guidance also states that multiple unincorporated businesses can be treated as one for purposes of that tax. [21]</p><p>Two separate LLC registrations therefore don’t guarantee two completely independent city tax outcomes.</p><p>If the software company develops trading infrastructure used by the trading company, I would also need to establish which entity owns that software and whether an intercompany license or services agreement is appropriate.</p><p>A holding company might eventually make sense if I own several operating businesses. A holding company is an entity that owns interests in other companies, often allowing ownership to be organized above them.</p><p>At the moment, another layer of ownership would create paperwork without an identified operational requirement. I would revisit it when actual financing, ownership, or organizational needs justify doing so.</p><h3>The Order I Would Follow</h3><p>After researching the formation process, I have a clearer idea of how I would proceed.</p><p>For the software company, I would begin with an inventory of the applications, code, domains, developer accounts, and other assets that might become company property. I would identify the agreements and revenue arrangements the business needs to support, compare a New York LLC with a corporation, and review the ongoing tax and administrative expenses.</p><p>My current preference is to evaluate a New York LLC first because I intend to operate an independently owned application business. That is a provisional choice, not a conclusion that an LLC is always preferable for software founders.</p><p>If I introduce outside investors, multiple equity owners, or employee stock arrangements, the corporation question becomes more significant. The Workiva example gives me confidence that a legal structure can evolve, while its detailed conversion disclosures remind me that changing ownership arrangements has real consequences.</p><p>For the trading project, I would continue building and testing the models before choosing an entity based on an imagined future trading business.</p><p>I would want reproducible research, independent evaluation, paper-trading results, and well-defined controls around order execution and financial exposure. Before live operation, I would also verify brokerage eligibility, regulatory requirements, taxes, and the capital available to absorb losses.</p><p>If the research doesn’t produce strategies that justify their costs and risks, I can continue developing the software without claiming to operate a profitable trading firm. That outcome would still provide useful engineering experience.</p><p>If the models demonstrate credible performance under appropriately controlled live conditions, I would have a much better basis for deciding what the trading company needs to own, how it should be financed, and which legal structure supports those activities.</p><p>The difference in timing comes from the responsibilities each business would assume. Existing applications already have identifiable software assets and publishing requirements. A prospective proprietary trading firm needs those organizational considerations too, but it also needs a demonstrated basis for putting capital at risk.</p><h3>Looking Back at the Gmail Invitation</h3><p>I still think about the ideas I had when Gmail was invitation-only.</p><p>An encrypted email service remains an interesting software challenge. So does building an operating system capable of attracting developers, device manufacturers, and users. Those ambitions were concerned with the technology itself and how many people might find it useful.</p><p>Working on applications has made me more interested in the decisions that determine whether a product can be supported over time. Someone has to own the code, maintain the service, honor agreements, account for the money, and accept responsibility when things go wrong.</p><p>The same questions become more consequential when software starts placing trades. The system needs trustworthy data, defensible research, reliable execution, and limits on the capital it can lose. A legal structure helps organize those responsibilities, but it doesn’t demonstrate that the trading model is any good.</p><p>The company histories I researched make more sense to me in that context. Apple and Microsoft started with partnerships. Google and Amazon incorporated early. Workiva operated as an LLC and became a corporation as its financing and ownership arrangements evolved. Their choices reflected different commercial circumstances, and none provides a universal blueprint.</p><p>I want to build a software company around applications that already exist, and I want to continue developing quantitative trading software toward a possible proprietary trading business. I can pursue both ambitions without pretending that they are equally ready for incorporation.</p><p>The next useful decision is identifying what each business actually needs to own, what it must be able to prove, and which responsibilities I am prepared to take on. That is a more concrete foundation for establishing companies than trying to reproduce the paperwork used by a successful founder decades ago.</p><h3>References</h3><p>[1] WIRED. “April 1, 2004: Gmail Hits Webmail G-Spot.” April 1, 2011. <a href="https://www.wired.com/2011/04/0401gmail-launches/">https://www.wired.com/2011/04/0401gmail-launches/</a></p><p>[2] GNU Privacy Guard. <em>GnuPG Documentation.</em> <a href="https://gnupg.org/documentation/manuals/gnupg/">https://gnupg.org/documentation/manuals/gnupg/</a></p><p>[3] Whitten, Alma, and J. D. Tygar. “Why Johnny Can’t Encrypt: A Usability Evaluation of PGP 5.0.” <em>8th USENIX Security Symposium</em>, 1999. <a href="https://www.usenix.org/conference/8th-usenix-security-symposium/presentation/why-johnny-cant-encrypt-usability-evaluation-p">https://www.usenix.org/conference/8th-usenix-security-symposium/presentation/why-johnny-cant-encrypt-usability-evaluation-p</a></p><p>[4] Computer History Museum. <em>Apple Timeline.</em> <a href="https://computerhistory.org/apple-timeline/">https://computerhistory.org/apple-timeline/</a></p><p>[5] Microsoft Learn. “The History of Microsoft — 1975.” <a href="https://learn.microsoft.com/en-us/shows/history/history-of-microsoft-1975">https://learn.microsoft.com/en-us/shows/history/history-of-microsoft-1975</a></p><p>[6] Microsoft Investor Relations. <em>Frequently Asked Questions.</em> <a href="https://www.microsoft.com/en-us/investor/faq.aspx">https://www.microsoft.com/en-us/investor/faq.aspx</a></p><p>[7] Google Inc. <em>2004 Registration Statement.</em> U.S. Securities and Exchange Commission. <a href="https://www.sec.gov/Archives/edgar/data/1288776/000119312504134174/ds1a.htm">https://www.sec.gov/Archives/edgar/data/1288776/000119312504134174/ds1a.htm</a></p><p>[8] Amazon Investor Relations. <em>Frequently Asked Questions.</em> <a href="https://ir.aboutamazon.com/faqs/default.aspx">https://ir.aboutamazon.com/faqs/default.aspx</a></p><p>[9] Workiva Inc. <em>Annual Report on Form 10-K for the Year Ended December 31, 2014.</em> U.S. Securities and Exchange Commission. <a href="https://www.sec.gov/Archives/edgar/data/1445305/000144530515000028/workiva10-k12x31x2014.htm">https://www.sec.gov/Archives/edgar/data/1445305/000144530515000028/workiva10-k12x31x2014.htm</a></p><p>[10] Jane Street Group, LLC. <em>2022 SEC Filing.</em> U.S. Securities and Exchange Commission. <a href="https://www.sec.gov/Archives/edgar/data/1595888/000110465922070509/tm2217771d1_406ba.htm">https://www.sec.gov/Archives/edgar/data/1595888/000110465922070509/tm2217771d1_406ba.htm</a></p><p>[11] U.S. Small Business Administration. <em>Launch Your Business.</em> <a href="https://www.sba.gov/counseling/launch-your-business/">https://www.sba.gov/counseling/launch-your-business/</a></p><p>[12] Internal Revenue Service. <em>Single-Member Limited Liability Companies.</em> <a href="https://www.irs.gov/businesses/small-businesses-self-employed/single-member-limited-liability-companies">https://www.irs.gov/businesses/small-businesses-self-employed/single-member-limited-liability-companies</a></p><p>[13] New York Department of State. <em>Forming a Limited Liability Company in New York.</em> <a href="https://dos.ny.gov/forming-limited-liability-company-new-york">https://dos.ny.gov/forming-limited-liability-company-new-york</a></p><p>[14] New York Department of State. <em>Certificate of Incorporation for Domestic Business Corporation.</em> <a href="https://dos.ny.gov/certificate-incorporation-domestic-business-corporation">https://dos.ny.gov/certificate-incorporation-domestic-business-corporation</a></p><p>[15] New York Department of State. <em>Biennial Statements for Business Corporations and Limited Liability Companies.</em> <a href="https://dos.ny.gov/biennial-statements-business-corporations-and-limited-liability-companies">https://dos.ny.gov/biennial-statements-business-corporations-and-limited-liability-companies</a></p><p>[16] Internal Revenue Service. <em>Employer Identification Number.</em> <a href="https://www.irs.gov/businesses/employer-identification-number">https://www.irs.gov/businesses/employer-identification-number</a></p><p>[17] Apple Developer. <em>Program Enrollment.</em> <a href="https://developer.apple.com/help/account/membership/program-enrollment">https://developer.apple.com/help/account/membership/program-enrollment</a></p><p>[18] Apple Developer. <em>Overview of App Transfer.</em> <a href="https://developer.apple.com/help/app-store-connect/transfer-an-app/overview-of-app-transfer/">https://developer.apple.com/help/app-store-connect/transfer-an-app/overview-of-app-transfer/</a></p><p>[19] New York Department of Taxation and Finance. <em>Computer Software — Tax Bulletin ST-128.</em> Updated October 16, 2026. <a href="https://www.tax.ny.gov/pubs_and_bulls/tg_bulletins/st/computer_software.htm">https://www.tax.ny.gov/pubs_and_bulls/tg_bulletins/st/computer_software.htm</a></p><p>[20] New York Department of Taxation and Finance. <em>Partnership, LLC, and LLP Annual Filing Fee.</em> <a href="https://www.tax.ny.gov/pit/efile/annual_filing_fee.htm">https://www.tax.ny.gov/pit/efile/annual_filing_fee.htm</a></p><p>[21] New York City Department of Finance. <em>Unincorporated Business Tax.</em> <a href="https://www.nyc.gov/site/finance/business/business-forms-unincorporated-business-tax-ubt.page">https://www.nyc.gov/site/finance/business/business-forms-unincorporated-business-tax-ubt.page</a></p><p>[22] Delaware General Assembly. <em>House Bill 400, 153rd General Assembly.</em> 2026. <a href="https://legis.delaware.gov/BillDetail/143069">https://legis.delaware.gov/BillDetail/143069</a></p><p>[23] Financial Crimes Enforcement Network. “FinCEN Removes Beneficial Ownership Reporting Requirements for U.S. Companies and U.S. Persons.” March 21, 2025. <a href="https://www.fincen.gov/news/news-releases/fincen-removes-beneficial-ownership-reporting-requirements-us-companies-and-us">https://www.fincen.gov/news/news-releases/fincen-removes-beneficial-ownership-reporting-requirements-us-companies-and-us</a></p><p>[24] New York Department of State. <em>Beneficial Ownership Disclosure: Frequently Asked Questions.</em> <a href="https://dos.ny.gov/beneficial-ownership-disclosure-frequently-asked-questions">https://dos.ny.gov/beneficial-ownership-disclosure-frequently-asked-questions</a></p><p>[25] U.S. Securities and Exchange Commission. <em>Guide to Broker-Dealer Registration.</em> <a href="https://www.sec.gov/divisions/marketreg/bdguide.htm">https://www.sec.gov/divisions/marketreg/bdguide.htm</a></p><p>[26] Investor.gov. <em>Investment Advisers.</em> U.S. Securities and Exchange Commission. <a href="https://www.investor.gov/introduction-investing/getting-started/working-investment-professional/investment-advisers">https://www.investor.gov/introduction-investing/getting-started/working-investment-professional/investment-advisers</a></p><p>[27] National Futures Association. <em>Commodity Trading Advisor Registration.</em> <a href="https://www.nfa.futures.org/registration-membership/who-has-to-register/cta.html">https://www.nfa.futures.org/registration-membership/who-has-to-register/cta.html</a></p><p>[28] National Futures Association. <em>Commodity Pool Operator Registration.</em> <a href="https://www.nfa.futures.org/registration-membership/who-has-to-register/cpo.html">https://www.nfa.futures.org/registration-membership/who-has-to-register/cpo.html</a></p><p>[29] U.S. Securities and Exchange Commission. “SEC Adopts Large Trader Reporting Regime.” July 26, 2011. <a href="https://www.sec.gov/newsroom/press-releases/2011-154-sec-adopts-large-trader-reporting-regime">https://www.sec.gov/newsroom/press-releases/2011-154-sec-adopts-large-trader-reporting-regime</a></p><p>[30] Financial Industry Regulatory Authority. <em>Regulatory Notice 15–09: Supervision and Control Practices for Algorithmic Trading Strategies.</em> March 2015. <a href="https://www.finra.org/sites/default/files/notice_doc_file_ref/Notice_Regulatory_15-09.pdf">https://www.finra.org/sites/default/files/notice_doc_file_ref/Notice_Regulatory_15-09.pdf</a></p><p>[31] Internal Revenue Service. <em>Topic №429: Traders in Securities.</em> <a href="https://www.irs.gov/taxtopics/tc429">https://www.irs.gov/taxtopics/tc429</a></p><p>[32] Investor.gov. <em>Investor Bulletin: Performance Claims.</em> September 15, 2022. <a href="https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-bulletins-47">https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-bulletins-47</a></p><p>[33] New York City Department of Small Business Services. <em>Get Free Legal Advice.</em> <a href="https://nyc-business.nyc.gov/nycbusiness/business-services/legal-assistance/get-free-legal-advice">https://nyc-business.nyc.gov/nycbusiness/business-services/legal-assistance/get-free-legal-advice</a></p><p>[34] Apple Developer. <em>Updating Your Account Information.</em> <a href="https://developer.apple.com/help/account/membership/updating-your-account-information/">https://developer.apple.com/help/account/membership/updating-your-account-information/</a></p><p>[35] Internal Revenue Service. <em>Instructions for Form 2553: Election by a Small Business Corporation.</em> <a href="https://www.irs.gov/instructions/i2553">https://www.irs.gov/instructions/i2553</a></p><p><em>Research current through October 21, 2026. The discussion of formation, taxation, and financial-market regulation is educational and reflects the author’s research and proposed plans. Specific legal and tax obligations require verification against the business’s actual activities.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=989184321cb9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learning to Read What macOS Is Telling You]]></title>
            <link>https://medium.com/@raell.dottin/learning-to-read-what-macos-is-telling-you-69f62aa2639e?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/69f62aa2639e</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 17:01:03 GMT</pubDate>
            <atom:updated>2026-11-09T17:01:03.062Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>What Howard Oakley’s writing taught me about Unified Logging, and why a Slack screen-sharing error is a useful place to put those lessons to work.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_kTLuaiDDl3g-DUt" /><figcaption>Photo by <a href="https://unsplash.com/@mihaiteslariu0?utm_source=medium&amp;utm_medium=referral">Teslariu Mihai</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>On October 14, Rich Trouton published a short article about determining how far back macOS Unified Logging records go. His question was familiar: if someone reports a failure from last week, can you still retrieve the events from that day? Trouton shows how to check the oldest and newest messages currently available in the log store. [3]</p><p>It reminded me of when I first encountered Unified Logging. I can’t identify the first article I read, but I’m fairly certain Howard Oakley at <em>The Eclectic Light Company</em> taught me most of what I needed to begin using it. His March 2018 article explains how recognizable events can serve as waypoints when navigating a large log, and his October 2018 article demonstrates how to isolate a reproducible problem in a short period of activity. [1] [2] I have continued to use the approach in my work with Macs, even as Apple has changed the details of the operating system.</p><p>What stayed with me was a method. A logging system becomes useful when you can connect an action to the processes that handled it and the outcome that followed. That connection can support a diagnosis; a matching error message, on its own, often cannot. Trouton’s article supplies a related caution: sometimes the evidence you need has already disappeared. I’ll explain both ideas using a documented Slack screen-sharing problem and the commands built into macOS.</p><h3>What Unified Logging actually records</h3><p>Apple introduced Unified Logging with macOS Sierra in 2016, replacing much of the old reliance on individual, human-readable text files. Applications and operating-system components now contribute structured diagnostic messages to a common system. A background process named logd helps handle those messages, storing some temporarily in memory and others on disk. Apple&#39;s Console application provides a graphical way to inspect them; Terminal provides the log command. [4]</p><p>A log entry can identify when something happened, the process that produced the message, a subsystem identifying its functional area, and some text describing the event. Those details are useful, but the volume can be daunting. An application opening a window may trigger messages from several supporting processes. If you search the entire store for an application name, you can find plenty of activity without finding an explanation.</p><p>Oakley’s 2018 troubleshooting article offered a much tighter procedure: reproduce the action, note the time, and examine a narrow interval on either side. He originally demonstrated it with his graphical tool Consolation. We can apply the same method with the log command in Terminal, using a screen-sharing failure to examine a privacy system that normally stays out of view. [2]</p><h3>When Slack says it can’t share a screen</h3><p>Slack documented a bug in which some people switched from its directly downloaded Mac application to the Mac App Store version, then found they couldn’t share their screens in huddles. Slack asked for screen-recording permission even though System Settings already showed the application as allowed. Toggling the setting didn’t solve it. Slack’s June 17, 2024 release notes acknowledge the problem on macOS Monterey and later. [5] [6]</p><p>This is a historical, documented failure, not a claim that today’s Slack releases all suffer from it. We also don’t have a saved logging trace from one of the affected Macs. We’ll use the case to build a repeatable investigation, and keep the difference between a plausible explanation and an established cause visible throughout.</p><p>The Mac setting involved is in System Settings → Privacy &amp; Security → Screen &amp; System Audio Recording. [7] The mechanism behind many macOS privacy decisions is called <strong>Transparency, Consent, and Control</strong>, or <strong>TCC</strong>. A background service named tccd helps evaluate requests to use protected capabilities. TCC maintains privacy decisions associated with applications, while Unified Logging can preserve messages about individual authorization requests. They are different kinds of records: a current permission setting isn&#39;t a transcript of the earlier attempt that failed.</p><p>Internally, the service associated with screen-capture authorization is called kTCCServiceScreenCapture. Applications may also use helper processes, so the executable doing a particular piece of work isn&#39;t necessarily the application whose authorization macOS evaluates. TCC uses an <strong>attribution chain</strong> to associate the request with the responsible application. Oakley&#39;s April 2026 investigation of protected-folder access includes real attribution-chain messages and authorization results. The experiment involved files rather than Slack, but it shows the kind of information a TCC investigation can reveal. [8]</p><p>For a confident authorization diagnosis, I would want four links: the moment someone started screen sharing, the corresponding protected-service request, the application TCC regarded as responsible, and a recorded decision consistent with the observed failure. If a denial belongs to a different process or occurred several minutes earlier, it cannot simply be assigned to this attempt. This is the reasoning the Terminal exercise needs to test, rather than a search for familiar words.</p><h3>Reconstructing the screen-sharing attempt</h3><p>Start by documenting the macOS version, Slack version, and whether the application came from the App Store or Slack’s website. Record the time of the failed attempt, preferably to the nearest second; running date in Terminal immediately beforehand is a simple way to capture a clock reading. Installation source is important in this case because Slack tied the historical issue to switching between versions. [5]</p><p>If the failure just occurred, open Terminal and search recent activity:</p><pre>/usr/bin/log show \<br>  --last 10m \<br>  --info \<br>  --style compact \<br>  --predicate &#39;subsystem == &quot;com.apple.TCC&quot; OR process == &quot;tccd&quot; OR process CONTAINS[c] &quot;Slack&quot;&#39;</pre><p>log show reads recorded messages. --last 10m limits the search to ten minutes; --info includes informational messages if they&#39;re still available; and --style compact makes the output easier to scan. A <strong>predicate</strong> is a filter. Here we&#39;re requesting messages from the TCC subsystem, the tccd process, or processes whose names contain Slack. The OR means any of those conditions is enough to match. These options are documented in the macOS log manual. [9]</p><p>Start with this broad filter rather than assuming every useful message will mention the screen-capture service. Look near the recorded failure time for a sequence of actions and authorization activity. Once you know what’s present, narrow the search to messages that explicitly identify screen-capture checks:</p><pre>/usr/bin/log show \<br>  --last 10m \<br>  --info \<br>  --style compact \<br>  --predicate &#39;eventMessage CONTAINS[c] &quot;kTCCServiceScreenCapture&quot;&#39;</pre><p>eventMessage is the message text, and CONTAINS[c] checks for a matching phrase without requiring identical capitalization. This query can narrow the results, but a matching service name isn&#39;t enough to establish that Slack&#39;s screen-sharing attempt was denied.</p><p>Oakley’s April 2026 investigation provides a particularly interesting example of why the surrounding context is important. During an experiment involving an Open and Save dialog, he encountered kTCCServiceScreenCapture in the logging activity, despite investigating file selection rather than screen capture. He repeated the experiment and observed the same unexpected behavior. [8]</p><p>The event was genuine, but identifying its purpose from the service name alone would have led to the wrong interpretation of the user’s action.</p><p>For our Slack investigation, we would need to establish the event’s timestamp, the application responsible for the request, and the authorization result. We also need to determine whether the activity belongs to the screen-sharing attempt rather than some unrelated action happening at approximately the same time.</p><p>If the historical query gives us little to work with, we can watch a new attempt as it happens:</p><pre>/usr/bin/log stream \<br>  --level info \<br>  --style compact \<br>  --predicate &#39;subsystem == &quot;com.apple.TCC&quot; OR process == &quot;tccd&quot; OR process CONTAINS[c] &quot;Slack&quot;&#39;</pre><p>log stream shows events as macOS produces them, rather than asking for older entries. Leave the command running, try to share the screen in Slack, and return to Terminal to stop collection with Control-C.</p><p>Watching the event live can expose informational activity that might not remain available for later queries. Apple’s logging documentation explains that the message level and configuration affect whether information is kept. [10]</p><h3>Interpreting the evidence</h3><p>Oakley’s published protected-folder experiment provides an example of how to read the sequence rather than an isolated line.</p><p>The recorded events identify the requesting service and the application involved through an AttributionChain. In one denied-access attempt, an AUTHREQ_RESULT entry is followed by another message explaining that the protected-folder request was denied. That lets him connect the attempted operation to an authorization outcome and to the application&#39;s inability to read the folder. [8]</p><p>For Slack, a denial tied to the responsible application and to the screen-sharing attempt would support an authorization-related explanation. We’d still investigate why macOS reached that decision.</p><p>If the log instead shows an applicable authorization succeeded, we’d look beyond the initial TCC check for capture-related failures. A successful authorization is not proof that all the subsequent work needed for screen sharing completed successfully.</p><p>If there is no matching TCC event, we’d first check whether our time window, predicate, and logging level were sufficient. Absence from a filtered log cannot prove the request never occurred.</p><p>Each observation changes the next step, but none should be stretched into a complete root cause without the corresponding evidence.</p><p>That limitation becomes important when comparing attempts before and after a repair. A successful share proves that the feature worked on the second attempt. It doesn’t automatically prove which internal state was wrong before the reset. Likewise, a logged TCC approval linked to the relevant request could narrow our search toward a later capture failure, but it would not guarantee that every prerequisite for screen sharing had been satisfied.</p><h3>Preserve the record, then repair</h3><p>If someone is trying to finish a meeting, a quick repair may reasonably take priority over an elaborate investigation. For a recurring failure, or one affecting many people in an organization, preserving the first useful trace can prevent repeated guesswork.</p><p>I would collect the available entries before changing the relevant privacy decision when that is practical:</p><pre>/usr/bin/log collect \<br>  --last 15m \<br>  --output &quot;$HOME/Desktop/slack-screen-sharing.logarchive&quot;</pre><p>This creates a .logarchive package containing available diagnostics from the selected interval. Collection may require administrator privileges.</p><p>The archive can be opened in Console or searched in Terminal. For example, this retrieves available TCC messages from the saved package:</p><pre>/usr/bin/log show \<br>  --archive &quot;$HOME/Desktop/slack-screen-sharing.logarchive&quot; \<br>  --info \<br>  --style compact \<br>  --predicate &#39;subsystem == &quot;com.apple.TCC&quot; OR process == &quot;tccd&quot;&#39;</pre><p>Treat the archive as sensitive diagnostic material, since system logs may expose details about application activity. A collected archive cannot recover messages already discarded. [9]</p><p>Slack’s support article documents the following reset for the historical bug:</p><pre>tccutil reset ScreenCapture com.tinyspeck.slackmacgap</pre><p>tccutil is Apple&#39;s command for managing stored privacy decisions. The command clears the screen-capture decision associated with the specified Slack application identifier. It doesn&#39;t quietly grant access; the user has to respond to a fresh prompt. Slack instructs affected users to quit and reopen its application before trying screen sharing again. [5] [11]</p><p>I would verify the installed application and applicability of that guidance before running the reset. Then I’d reproduce the action again and compare the new log window with the one I saved.</p><p>If the feature works, I can record an effective repair. If the earlier messages show a corresponding difference in authorization, I may have evidence for a more specific explanation. Without that earlier trace, it would be overstating the result to assert precisely what happened inside TCC.</p><h3>The limits of log retention</h3><p>The limits of what we can prove bring us back to Trouton. Someone might report the same problem days after it happened, when there is no longer an opportunity to capture the original request.</p><p>Before reading hundreds of messages from the wrong period, ask the logging system what it still holds:</p><pre>/usr/bin/log stats --overview</pre><p>Trouton’s October 2026 article explains that the overview includes start and end values for the oldest and newest stored events. The command can take several minutes to calculate the result. It is a useful way to establish whether the available history might extend back to the incident date. [3]</p><p>The output can also be requested as JavaScript Object Notation (JSON), a structured text format commonly used to exchange data between programs:</p><pre>/usr/bin/log stats --overview --style json</pre><p>Trouton demonstrates using this format to extract the time boundaries programmatically. That can be useful when collecting diagnostics across multiple computers, although the timestamps alone still require careful interpretation. [3]</p><p>A start value predating the failure doesn&#39;t guarantee that relevant events survived.</p><p>Apple’s Unified Logging system manages stored messages according to their type, volume, and space requirements. In one of Oakley’s September 2025 measurements, an initial sample of 22,783 entries had fallen to 18,655 after five minutes; a substantial portion of that sample was never retained on disk. That observation describes his measured workload, not a universal percentage for every Mac. [12]</p><p>Records that survive the first few minutes can be removed later. Oakley documented older periods with isolated surviving messages and substantial gaps in between. A Mac might therefore have an entry from last month while having no relevant records at all from the day the user reports. [13]</p><p>The storage system also treats different categories of records differently. Ordinary retained messages commonly reside in the Persist area of the logging store, while the Special area can contain certain events that survive after much of the surrounding activity has disappeared.</p><p>Oakley’s measurements of these storage areas help explain why retention cannot be reduced to a fixed number of days. The rate at which a Mac produces messages and the categories being stored both influence what remains available. [12] [13]</p><p>That design makes more sense when you remember what Unified Logging is built to do. It prioritizes efficient diagnostics and controlled storage, with privacy protections and differing message levels. It isn’t a guaranteed complete audit trail. Apple may also replace sensitive values in messages with &lt;private&gt;, limiting what an investigator can inspect even when an event survives. [4] [10]</p><p>For an active, reproducible issue, this argues for collecting evidence promptly. For an old issue, it argues for modesty about conclusions drawn from missing events.</p><p>If the records don’t cover the failure, the answer may be that the original authorization decision is no longer recoverable from this Mac’s log store. A retained timestamp gives the investigation a possible boundary, not a promise of complete coverage.</p><h3>The lesson that has stayed useful</h3><p>I can’t point to the precise Oakley article I first read. What I recognize across his early writing is the discipline of anchoring a search to a real action and using surrounding events to establish what macOS did. Trouton’s recent retention article adds a practical check on whether the record is still available. [1] [2] [3]</p><p>For the Slack case, that method means recording the failed share, investigating the responsible process and any authorization result, preserving the available messages, and testing the documented repair.</p><p>The result might identify a TCC denial, indicate that capture failed after authorization, or show that the relevant evidence was never retained. Those outcomes lead to different next steps, and knowing which one the log supports is more useful than simply finding a familiar error string.</p><p>That’s the part of Unified Logging I want to pass on. When a Mac behaves unexpectedly, its logs can help explain what happened, provided we ask a narrow enough question and remain careful about what the surviving messages actually show.</p><h3>References</h3><p>[1] Howard Oakley. “macOS Unified log: 3 finding your way.” <em>The Eclectic Light Company</em>, March 21, 2018.<br><a href="https://eclecticlight.co/2018/03/21/macos-unified-log-3-finding-your-way/">https://eclecticlight.co/2018/03/21/macos-unified-log-3-finding-your-way/</a></p><p>[2] Howard Oakley. “How to use the unified log to see what’s going wrong.” <em>The Eclectic Light Company</em>, October 12, 2018.<br><a href="https://eclecticlight.co/2018/10/12/how-to-use-the-unified-log-to-see-whats-going-wrong/">https://eclecticlight.co/2018/10/12/how-to-use-the-unified-log-to-see-whats-going-wrong/</a></p><p>[3] Rich Trouton. “Determining log retention periods for macOS unified logging.” <em>Der Flounder</em>, October 14, 2026.<br><a href="https://derflounder.wordpress.com/2026/10/14/determining-log-retention-periods-for-macos-unified-logging/">https://derflounder.wordpress.com/2026/10/14/determining-log-retention-periods-for-macos-unified-logging/</a></p><p>[4] Howard Oakley. “Inside the Unified Log 1: Goals and architecture.” <em>The Eclectic Light Company</em>, September 23, 2025.<br><a href="https://eclecticlight.co/2025/09/23/inside-the-unified-log-1-goals-and-architecture/">https://eclecticlight.co/2025/09/23/inside-the-unified-log-1-goals-and-architecture/</a></p><p>[5] Slack. “Troubleshoot huddles screen sharing bug on the Slack Mac desktop app.” <em>Slack Help Center</em>.<br><a href="https://slack.com/help/articles/29407960918291-Troubleshoot-huddles-screen-sharing-bug-on-the-Slack-Mac-desktop-app">https://slack.com/help/articles/29407960918291-Troubleshoot-huddles-screen-sharing-bug-on-the-Slack-Mac-desktop-app</a></p><p>[6] Slack. “Slack for Mac — Release Notes.” Version 4.39.88, June 17, 2024.<br><a href="https://slack.com/release-notes/mac">https://slack.com/release-notes/mac</a></p><p>[7] Apple. “Control access to screen and system audio recording on Mac.” <em>Mac User Guide</em>.<br><a href="https://support.apple.com/guide/mac-help/mchld6aa7d23/mac">https://support.apple.com/guide/mac-help/mchld6aa7d23/mac</a></p><p>[8] Howard Oakley. “Privacy: Files &amp; Folders or Full Disk Access?” <em>The Eclectic Light Company</em>, April 8, 2026.<br><a href="https://eclecticlight.co/2026/04/08/privacy-files-folders-or-full-disk-access/">https://eclecticlight.co/2026/04/08/privacy-files-folders-or-full-disk-access/</a></p><p>[9] Apple. “log(1) — General Commands Manual.” macOS command-line reference.<br><a href="https://manp.gs/mac/1/log">https://manp.gs/mac/1/log</a></p><p>[10] Apple Developer Documentation. “Generating Log Messages from Your Code.”<br><a href="https://developer.apple.com/documentation/os/generating-log-messages-from-your-code">https://developer.apple.com/documentation/os/generating-log-messages-from-your-code</a></p><p>[11] Apple. “tccutil(1) — General Commands Manual.” macOS command-line reference.<br><a href="https://manp.gs/mac/1/tccutil">https://manp.gs/mac/1/tccutil</a></p><p>[12] Howard Oakley. “Inside the Unified Log 3: Log storage and attrition.” <em>The Eclectic Light Company</em>, September 29, 2025.<br><a href="https://eclecticlight.co/2025/09/29/inside-the-unified-log-3-log-storage-and-attrition/">https://eclecticlight.co/2025/09/29/inside-the-unified-log-3-log-storage-and-attrition/</a></p><p>[13] Howard Oakley. “How long does the log keep entries?” <em>The Eclectic Light Company</em>, March 12, 2026.<br><a href="https://eclecticlight.co/2026/03/12/how-long-does-the-log-keep-entries/">https://eclecticlight.co/2026/03/12/how-long-does-the-log-keep-entries/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69f62aa2639e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hermes Agent and the Mythology of Agile Project Management]]></title>
            <link>https://medium.com/@raell.dottin/hermes-agent-and-the-mythology-of-agile-project-management-d36513c90152?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/d36513c90152</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[project-management]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 17:01:01 GMT</pubDate>
            <atom:updated>2026-11-08T17:01:01.765Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>What happens to software development’s familiar rituals when AI agents can manage the board and perform the work?</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2f7LFAyxzm-H5qV2" /><figcaption>Photo by <a href="https://unsplash.com/@silversurfer79?utm_source=medium&amp;utm_medium=referral">Adrian Greaves</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>A Board That Can Do the Work</h3><p>When I started using Hermes Kanban, I expected the biggest advantage to come from having several AI agents working on a project simultaneously. I had already been using Hermes to write code, investigate problems, and run tests. Giving it a board where tasks could be distributed among specialized workers seemed like a reasonable next step.</p><p>The arrangement I eventually settled on was more conservative than I expected. I generally keep one agent responsible for implementation and assign other agents to investigate the ways that implementation might be wrong.</p><p>For Opnory, an AI-powered IT service desk, that means examining architectural boundaries, authorization, and security. For Alepes, a rules-based financial system, the concerns include exact-money arithmetic, transaction ordering, reconciliation, and whether changes preserve the underlying financial rules.</p><p>The other agents are there because I want different kinds of scrutiny applied to the same work.</p><p>Hermes Kanban provides a persistent record of these assignments. A worker can receive a task, perform its work, request a review, and record the result. Other agents can examine that work, while the system manages dependencies, execution attempts, and failures [2].</p><p>This changes the relationship between a project board and the work it describes. The board can initiate execution rather than waiting for someone to read a ticket and begin working.</p><p>I’ve spent much of my career supporting the systems that technology organizations depend on. Project-management tools are familiar territory, and I understand why organizations use them. They establish priorities, coordinate people with different responsibilities, and provide visibility into complicated projects.</p><p>I’ve also seen how easily the administrative process surrounding a project can consume considerable time.</p><p>A team updates its tickets for a status meeting. The meeting produces information for another report. That report feeds a dashboard, which raises questions that require somebody to update the tickets again.</p><p>Some of this coordination is necessary, particularly in larger organizations. But I’ve encountered processes where maintaining a description of the work becomes a substantial activity of its own.</p><p>Hermes presents an opportunity to examine those practices because software agents can perform some of the coordination while carrying out the assignments.</p><p>There are two questions worth separating. The first concerns how much administrative coordination can be automated. The second concerns whether doing so improves the delivery of useful software.</p><p>An automated system may answer the first question convincingly while performing poorly on the second. Understanding why requires looking at the original intentions of Agile and Kanban, along with what recent research tells us about AI-assisted development.</p><h3>What Hermes Actually Automates</h3><p>Hermes Agent, developed by Nous Research, is a framework for running AI agents that use tools, execute commands, maintain context, and complete multi-step assignments.</p><p>Its Kanban system adds persistent task management. Persistent means that assignments and their histories remain recorded even when an individual agent process stops or the system restarts.</p><p>The underlying implementation uses SQLite, a lightweight database, to store information about tasks, execution attempts, dependencies, and handoffs. A dispatcher monitors the board and starts workers when their assignments are eligible to run [2].</p><p>An orchestrator is an agent responsible for organizing a larger objective. It can divide an assignment into smaller tasks and establish dependencies between them. A worker agent then performs the assigned task using its configured AI model, instructions, and tools.</p><p>Consider a request to improve automated testing in an application. An orchestrator could assign one worker to investigate gaps in test coverage and another to implement the necessary changes. A reviewer could then evaluate the implementation before the system accepts the result.</p><p>Some assignments can run independently, while others need to wait until their prerequisites are complete. Hermes records those relationships and makes the resulting execution history available through its command-line interface, agent tools, and graphical dashboard [2][8].</p><p>The Kanban dashboard plugin is the visual interface. The dispatcher, task database, and agent tools provide the underlying coordination. The distinction is important because work can continue without somebody keeping the dashboard open.</p><p>Hermes also supports configurable execution limits, reviews, and recovery mechanisms. Workers can report that they are blocked, request changes, or record a completion summary. A task that repeatedly fails can stop after exhausting its permitted attempts [2][9].</p><p>These capabilities address problems that become increasingly important when software is running other software.</p><p>An AI worker can misunderstand a requirement, encounter an unavailable service, or exhaust its execution budget. A long-running task needs a predictable response to those conditions, particularly if its operator isn’t watching every step.</p><p>I’ve seen the value of that persistence in my own Alepes workflow. During one repository audit, a provider quota error caused the assignment to be requeued rather than counted as a failed implementation attempt. A subsequent run completed the audit. The task history preserved both events, allowing me to examine the recovery rather than reconstruct it from an interrupted conversation.</p><p>The successful recovery established that the workflow could continue after that particular failure. It didn’t independently establish that every conclusion in the resulting audit was correct.</p><p>That separation becomes central to the discussion of automated project management.</p><p>Hermes can make task execution more autonomous and its history easier to inspect. Whether the resulting work is useful depends on requirements, verification, and decisions that sit beyond the mechanics of moving an assignment through the board.</p><h3>Agile Wasn’t Supposed to Become This Complicated</h3><p>In February 2001, seventeen software practitioners met in Snowbird, Utah, and produced the <em>Manifesto for Agile Software Development</em>.</p><p>The document established four priorities: individuals and interactions, working software, customer collaboration, and responding to change [1].</p><p>Its authors were reacting to development practices that could become overly focused on plans, documentation, and formal procedures before customers had an opportunity to evaluate functioning software.</p><p>The manifesto didn’t propose eliminating planning or documentation. It acknowledged their usefulness while arguing for greater emphasis on the activities that helped teams deliver and adapt.</p><p>The accompanying principles encouraged frequent delivery, technical excellence, sustainable development, and regular examination of how teams could improve their work [3].</p><p>These ideas addressed genuine engineering problems. Projects could spend months defining requirements, preparing plans, and negotiating responsibilities before discovering that the resulting software didn’t adequately address the customer’s needs.</p><p>Agile emphasized shorter feedback cycles and closer collaboration among the people responsible for making decisions and implementing them.</p><p>Over time, organizations adopted frameworks intended to put these principles into practice.</p><p>Scrum organizes development through defined responsibilities, recurring events, and fixed-length periods called sprints. The official Scrum Guide describes a lightweight framework built around transparency, inspection, and adaptation [4].</p><p>Kanban approaches the problem through the flow of work rather than a prescribed development cycle [5].</p><p>The two approaches can coexist, but they aren’t interchangeable. Neither should be confused with the entire philosophy of Agile.</p><p>That distinction becomes important when organizations begin treating particular administrative practices as requirements of Agile itself.</p><p>The original manifesto didn’t prescribe two-week sprints, story points, or a specific ticket-management application. Even Scrum’s official guide doesn’t require story points, leaving teams to select suitable techniques for estimating their work [4].</p><p>Organizations introduced additional procedures for understandable reasons. Large engineering departments need to coordinate budgets, personnel, dependencies, and delivery commitments. Regulated businesses may need documented approvals and evidence that specific controls were followed.</p><p>Formal processes can help with those responsibilities.</p><p>The difficulty arises when compliance with the process becomes an adequate substitute for examining the result.</p><p>In 2018, Martin Fowler, one of the original Agile Manifesto signatories, criticized what he called the <em>Agile Industrial Complex</em>. Speaking at Agile Australia, Fowler described organizations imposing standardized development practices while undermining the team autonomy and technical excellence that Agile was intended to encourage [7].</p><p>Fowler wasn’t arguing that teams should abandon organization or accountability. His criticism concerned the tendency to prescribe processes without sufficiently examining whether they improved the development of software.</p><p>That concern becomes particularly relevant when an AI system can perform many of the administrative activities involved in coordinating tasks.</p><p>If the purpose of a recurring meeting is to communicate information already available in an execution log, automation may reduce the need for that meeting.</p><p>But a meeting in which people negotiate priorities, resolve conflicting requirements, or reconsider a product decision serves a different purpose. An automated status report cannot settle those questions merely by containing accurate information.</p><p>The useful test is to examine what a practice accomplishes and whether the same outcome can be achieved more effectively.</p><h3>Kanban Was About Flow Before It Was About Boards</h3><p>Kanban originated in manufacturing, particularly through Toyota’s efforts to coordinate production according to demand.</p><p>The Toyota Production System used signals to indicate when additional parts were needed. These practices supported just-in-time manufacturing, in which production and replenishment are aligned with demand rather than allowing unnecessary inventory to accumulate [6].</p><p>Software engineering later adapted the underlying principles.</p><p>David J. Anderson describes introducing a pull-based system at Microsoft in 2004. A pull system allows work to enter a stage when the resources responsible for that stage have capacity to handle it.</p><p>His early Microsoft implementation didn’t use a physical Kanban board. It was concerned with controlling how work entered the system and how that work progressed [16].</p><p>The familiar columns labeled To Do, In Progress, and Done came to represent this approach, but the visual board was never sufficient on its own.</p><p>The May 2025 <em>Kanban Guide</em> defines Kanban as a strategy for optimizing the flow of value through a process. It requires a defined workflow, active management of work, and continuous improvement [5].</p><p>One of its central concepts is <strong>work in progress (WIP)</strong>, meaning work that has started but hasn’t finished.</p><p>Imagine a team capable of completing five assignments each week. If management continually introduces new assignments faster than the team can finish them, the number of active tasks increases.</p><p>Everyone may remain busy, but unfinished work accumulates. Some assignments wait for information, others compete for attention, and the eventual completion of each item becomes harder to predict.</p><p>The Kanban Guide specifies four minimum flow measurements. Throughput counts the number of completed items during a period. Cycle time measures how long an item takes from starting to finishing. Work-item age measures how long an unfinished item has remained active. Work in progress records how many items are currently underway [5].</p><p>These measurements describe different aspects of the workflow.</p><p>Hermes provides configurable concurrency limits that control how many workers can run simultaneously. Its documented defaults leave both overall and per-profile concurrency unrestricted unless limits are supplied [2].</p><p>That is useful flexibility, although a concurrency limit isn’t necessarily equivalent to controlling work in progress across the entire delivery process.</p><p>A task may finish its agent execution and still be waiting for review, integration, deployment, or an external decision. Limiting active workers addresses one part of the workflow without necessarily controlling those downstream queues.</p><p>This becomes particularly interesting with AI agents because additional execution capacity can be relatively easy to obtain.</p><p>If I have several available models and enough computing resources, I can start more workers. But those workers may still depend on the same repository, test environment, reviewer, or deployment system.</p><p>I encountered a related problem when experimenting with multiple agents modifying the same codebase. One worker could change an interface while another was still reasoning about the previous version. What appeared to be an opportunity for parallel implementation introduced additional synchronization work.</p><p>That experience influenced the arrangement I now prefer: one implementation owner, with other agents performing targeted reviews.</p><p>The arrangement doesn’t eliminate coordination. It controls where coordination is necessary and reduces the likelihood that several workers will interfere with one another’s changes.</p><p>The original Kanban concern with unfinished work becomes more useful as the cost of starting additional assignments decreases.</p><p>If initiating work is inexpensive but completing and verifying it remains constrained, the system needs to pay close attention to how much unfinished work it is creating.</p><h3>The Work Gets Faster. The Review Queue Gets Longer.</h3><p>Recent research on AI-assisted software development offers useful evidence about what happens when the production of code accelerates.</p><p>The DevOps Research and Assessment program, commonly known as DORA, studies the engineering practices associated with software-delivery performance.</p><p>Its 2024 report found that greater AI adoption was associated with improvements in individual development activities but negative relationships with delivery throughput and stability [11].</p><p>Delivery throughput concerns how frequently software changes reach production. Stability concerns whether those changes behave reliably after release.</p><p>The results suggested that improvements in individual activities weren’t necessarily translating into improvements across the complete delivery process.</p><p>DORA’s 2025 report presented a somewhat more encouraging picture. It found a positive relationship between AI adoption and delivery throughput, while the negative relationship with delivery stability remained. The researchers characterized AI as an amplifier of the engineering capabilities and organizational weaknesses already present in a development environment [10].</p><p>These studies rely substantially on organizational surveys and observational analysis. They identify relationships rather than establishing that adopting AI will cause a particular result in every team.</p><p>A 2026 Harvard working paper provides more detailed evidence about where the constraints can appear.</p><p>In <em>Artificial Intelligence in the Firm: Bottlenecks in Software Production</em>, Fiona Chen and James Stratton examined approximately 300 million workplace events across 718 firms. Their dataset included repository activity, issue tracking, and organizational records [15].</p><p>The researchers used differences in when firms adopted AI coding tools to estimate how adoption affected software production.</p><p>For organizations adopting more autonomous coding agents, they estimated increases of approximately 30% in lines of code added, 20% in commits, and 23% in pull requests.</p><p>A commit records a change in a software repository. A pull request proposes incorporating code changes into a shared codebase, generally following testing and review.</p><p>Those figures indicate considerably more coding activity.</p><p>However, the researchers didn’t identify a statistically significant increase in resolved Jira issues or larger project groupings called epics.</p><p>That finding doesn’t prove that completed work failed to improve. It means the study didn’t establish an increase in those particular measures with sufficient statistical confidence.</p><p>The downstream findings are more informative.</p><p>Following AI-agent adoption, the researchers estimated a 49% increase in elapsed time between pull-request submission and merger. Comments per pull request increased by 35%, and the proportion of pull requests receiving formal requests for changes nearly doubled [15].</p><p>The 49% figure measures the elapsed submission-to-merge interval, including waiting time. It should not be interpreted as a direct measurement of the hours reviewers spent examining code.</p><p>The additional comments and revision requests provide separate evidence that the review process was experiencing greater demands.</p><p>There are several possible contributors. More proposed changes create more material to evaluate. Changes in the complexity or quality of submissions may also influence the amount of review required.</p><p>The researchers developed a model of software production examining how coding and code review interact. Their findings suggest that increasing coding productivity can create downstream constraints that limit improvements in completed output [15].</p><p>The paper is observational research, and its estimates depend on assumptions about how the studied firms would have performed without adoption. It isn’t a controlled experiment of Hermes.</p><p>Nevertheless, it describes a problem that is directly relevant to automated development workflows.</p><p>An organization can increase its ability to generate code without increasing its ability to approve, integrate, and release that code.</p><p>There is also experimental evidence that cautions against assuming productivity improvements.</p><p>In July 2025, Model Evaluation &amp; Threat Research (METR) published a randomized trial involving sixteen experienced open-source developers working on tasks in repositories they knew well.</p><p>The researchers assigned tasks to conditions that either permitted or prohibited AI assistance. Developers using the tools evaluated in that study took approximately 19% longer to complete their assignments [13].</p><p>The result was surprising, including to some participants who believed the tools had helped them.</p><p>The experiment provides evidence about a particular group of developers, repositories, and AI systems available during early 2025. It doesn’t establish a permanent relationship between AI assistance and developer productivity.</p><p>In February 2026, METR reported preliminary evidence that newer systems might provide greater speedups. However, selection effects and difficulties measuring work involving multiple agents made its updated estimates unreliable for determining the size of those improvements [14].</p><p>The findings have to be interpreted within their respective research designs.</p><p>DORA examines organizational patterns. METR investigates developer productivity through controlled experiments. Chen and Stratton analyze workplace activity and downstream production constraints.</p><p>None measures the precise performance of my Hermes configuration.</p><p>Taken together, they provide a reason to evaluate automated development across the complete workflow rather than treating code generation as the final outcome.</p><h3>Three Different Kinds of Capacity</h3><p>The research becomes easier to apply if we separate three forms of capacity within an engineering organization.</p><p>The first is <strong>execution capacity</strong>: the ability to produce proposed changes. It includes developers writing code, AI agents implementing assignments, and the tools that support those activities.</p><p>The second is <strong>verification capacity</strong>: the ability to establish whether those changes satisfy their requirements. Automated tests, code review, integration checks, security assessments, and other forms of quality control contribute to this capacity.</p><p>The third is <strong>decision-making capacity</strong>: the ability to establish priorities, define acceptable outcomes, resolve conflicts, and determine which risks an organization is willing to accept.</p><p>These categories are an analytical framework for understanding the workflow, not a formal model established by the cited studies. They overlap, and decisions can occur throughout development rather than only after verification.</p><p>They help explain why increasing the number of available agents doesn’t necessarily produce a comparable increase in delivered software.</p><p>Hermes can expand execution capacity by assigning work to several agents. It can also assist verification by running tests and delegating reviews. But those mechanisms depend on the quality of the instructions, the evidence available, and the authority assigned to each worker.</p><p>My experience with Opnory and Alepes illustrates why I began separating these responsibilities.</p><p>For Opnory, an implementation might satisfy the feature request while introducing a problem with tenant isolation, the rules that prevent one customer’s data or operations from crossing into another customer’s environment.</p><p>I want a security reviewer to examine that possibility independently of the worker responsible for implementation.</p><p>For Alepes, the concern might be a change that passes its immediate tests but violates a financial rule governing allocation, rounding, or reconciliation.</p><p>An agent assigned to financial invariants can examine those requirements while another searches for unusual inputs and failure conditions.</p><p>The additional reviewers increase the amount of verification work the system can perform concurrently. They also consume resources and can produce duplicate or inconclusive findings.</p><p>A separate verifier then has to determine whether the available evidence supports completion. That introduces another checkpoint, with its own cost and possible delays.</p><p>This arrangement is not necessarily faster than using a single agent for every assignment. For routine changes with straightforward acceptance criteria, the extra coordination may be difficult to justify.</p><p>For changes where an incorrect result could have significant consequences, I find the additional scrutiny more useful.</p><p>There is an important technical qualification. Assigning an agent a read-only review role through its instructions doesn’t automatically enforce filesystem permissions. If a review genuinely requires isolation from the implementation, that isolation needs to be established through the execution environment or access controls.</p><p>Similarly, a second AI agent doesn’t provide independence merely because it has a different name. Reviewers can share incorrect assumptions, misunderstand the same requirement, or accept convincing but inadequate evidence.</p><p>The relevant question is whether the review introduces a meaningful opportunity to discover failures that the implementation process would otherwise overlook.</p><p>This is where the original Agile emphasis on technical excellence and feedback becomes useful again [3].</p><h3>Who Decides That the Work Is Finished?</h3><p>Hermes includes a goal-oriented execution mode that allows a worker to continue through multiple attempts while an auxiliary AI evaluator checks progress against the task’s requirements.</p><p>When a task is created with --goal, its title and body provide the objective and acceptance criteria. The evaluator can determine whether the worker should continue, whether the goal appears to have been achieved, or whether the assignment cannot be completed as written [2].</p><p>Acceptance criteria are the conditions that must be satisfied before work can be considered successful.</p><p>For example, an assignment might require implementing a feature, adding automated tests, and verifying that the application builds successfully.</p><p>Those requirements are useful because they make the expected result explicit. They also determine what the automated evaluation can reasonably establish.</p><p>Consider a task involving the backtesting process for a trading model.</p><p>Backtesting means evaluating a strategy against historical market data to examine how it would have performed under the simulation’s assumptions.</p><p>An agent might implement new tests, correct an error in the historical-data pipeline, and produce a report showing that all configured tests passed.</p><p>That would be meaningful engineering progress.</p><p>It wouldn’t establish that the strategy is economically sound. Historical performance might depend on unrealistic transaction costs, assumptions about available liquidity, or choices that fit past data unusually well without generalizing to new conditions.</p><p>The implementation could satisfy its software requirements while the broader financial objective remained unproven.</p><p>This is a familiar problem in software development. A test suite can establish the behavior it was designed to examine, but requirements that were never tested remain outside its conclusions.</p><p>Hermes’s goal evaluator also has an operational limitation. Its documentation explains that if an evaluator request fails because of an unavailable service, authentication problem, or timeout, the system can allow a lifecycle handoff while recording a warning [2].</p><p>That behavior prevents a temporary evaluation failure from permanently obstructing human review. It also means operators shouldn’t treat every completed card as proof that an independent evaluator successfully examined the result.</p><p>Automated testing and continuous integration can provide additional evidence. <strong>Continuous integration (CI)</strong> refers to regularly combining software changes and checking them automatically so that problems can be identified before acceptance.</p><p>A successful build establishes that the software compiled under the tested conditions. Passing tests establish that the behaviors covered by those tests produced expected results. Reviewers can examine whether the implementation is maintainable, secure, compatible, and appropriate for the surrounding system.</p><p>These checks provide different forms of evidence.</p><p>Scrum addresses a related problem through its Definition of Done, a shared understanding of the quality requirements that completed work must satisfy [4].</p><p>There is a useful lesson here for AI-driven development.</p><p>The ability to automate implementation makes clear acceptance criteria increasingly valuable. An agent may be capable of spending substantial time pursuing an inadequate instruction, producing plausible results while missing a requirement that was never expressed.</p><p>At the same time, verification has a cost. Applying extensive independent review to every trivial change can produce overhead without a corresponding benefit.</p><p>The appropriate level of scrutiny depends on the consequences of failure and the evidence needed to justify accepting the work.</p><h3>Which Agile Practices Still Help?</h3><p>An automated development workflow still requires project management, but it changes which activities need direct human participation.</p><p>Status reporting is an obvious opportunity for automation. Hermes records task assignments, execution attempts, comments, failures, and completion summaries. Operators can inspect that history without asking a worker to reconstruct it from memory.</p><p>That information can reduce the effort needed to understand what happened.</p><p>Planning and adaptation are more complicated.</p><p>Scrum’s Daily Scrum, for example, is intended to help developers inspect progress toward their goal and adapt their plan. Treating it only as a meeting where people report what they did yesterday misunderstands its stated purpose [4].</p><p>If a team uses the meeting primarily to exchange information already captured by its tools, the format may deserve reconsideration. If the meeting helps people resolve conflicting priorities or identify dependencies that the tools cannot determine, it can remain valuable.</p><p>The same applies to other organizational procedures.</p><p>A release approval that exists solely to repeat automated test results may add little information. An approval in which a responsible person evaluates operational risk or determines whether a regulatory requirement has been satisfied serves a more substantial purpose.</p><p>Autonomous agents may even increase the need for some forms of coordination.</p><p>Multiple workers can modify shared components, compete for test environments, make incompatible assumptions, or produce changes that are individually reasonable but collectively difficult to integrate.</p><p>Explicit ownership, dependency management, and review checkpoints can prevent these problems. In that setting, a disciplined process enables autonomy rather than obstructing it.</p><p>Kanban provides a useful way to examine the resulting workflow because it directs attention toward unfinished work and the conditions preventing its completion [5].</p><p>The practical question is whether a particular activity helps people make decisions, manage risk, improve quality, or deliver useful results.</p><p>If it accomplishes one of those purposes, automation may change how the activity is performed without removing the need for it.</p><p>If it primarily duplicates information that already exists elsewhere, the organization has a reasonable opportunity to simplify the process.</p><p>The original Agile principles encourage teams to examine their practices and adjust them based on what they learn [3].</p><p>That recommendation remains applicable when some members of the development process are AI agents.</p><h3>Measuring the Result</h3><p>If I want to establish whether Hermes improves productivity, I need measurements that account for completed and accepted work.</p><p>Kanban’s flow measurements provide a useful foundation. Work in progress reveals accumulating assignments. Cycle time shows how long tasks take to finish, while work-item age identifies assignments that may be stalled [5].</p><p>For AI-agent workflows, I would also examine how often proposed changes pass review, how much rework they require, and how frequently an operator has to intervene.</p><p>Model and infrastructure costs belong in the same analysis.</p><p>A worker that produces an implementation in ten minutes may appear faster than one that takes thirty. If the first result requires several review cycles, additional model calls, and extensive corrections, the initial advantage may disappear.</p><p>Likewise, an expensive model may be cost-effective for a difficult assignment if it avoids repeated failures. A less expensive model may be entirely adequate for routine work with clear requirements.</p><p>The appropriate comparison is the cost and time required to produce an acceptable result under equivalent quality requirements.</p><p>I would also distinguish completed assignments from results that have survived integration and deployment. An implementation can pass review and still encounter problems when introduced into the larger system.</p><p>These measurements could support a controlled comparison between a conventional supervised workflow and one coordinated through Hermes.</p><p>The tasks would need comparable complexity and acceptance criteria, with repeated observations rather than one impressive demonstration. The comparison should include review effort and failures discovered after completion, not simply implementation time.</p><p>Another useful experiment would vary the number of concurrent agents while keeping review capacity constant.</p><p>That would help determine whether additional execution capacity increases accepted output or primarily creates more work waiting to be examined.</p><p>The purpose of measuring these outcomes is to make better decisions about the workflow. The results may justify additional agents, stricter concurrency limits, different reviewers, or a return to a simpler process for particular kinds of work.</p><p>That is more useful than assuming a larger swarm must be a more capable one.</p><h3>Back to the Board</h3><p>My original reason for adopting Hermes Kanban was practical. I wanted development assignments to continue without requiring me to supervise every step, and I wanted a better record of how those assignments were completed.</p><p>The system has helped me organize work across projects with different technical risks. It has also made me more deliberate about separating implementation from verification.</p><p>I still use a single agent for many assignments. For changes involving security boundaries, financial rules, or complicated dependencies, I prefer giving additional agents specific reasons to examine the work.</p><p>That arrangement creates its own coordination costs, which means it needs to justify them.</p><p>The experience has changed how I think about project-management practices more generally.</p><p>An automated board can reduce the effort involved in assigning work, collecting status, and maintaining execution histories. Those capabilities make some administrative routines easier to question.</p><p>But the ability to generate proposed changes increases the importance of understanding which changes should be made, how they should be evaluated, and what evidence is sufficient to accept them.</p><p>The research on AI-assisted development suggests that these concerns extend beyond my own projects. Faster code production can place additional demands on review and integration, while organizational decisions continue to determine how the resulting software is used [10][15].</p><p>Agile began with an argument for adapting development practices to improve the delivery of useful software. Kanban offered a way to understand and improve how work moves through a system.</p><p>Both remain relevant when execution is increasingly automated.</p><p>The opportunity is to examine our processes with greater precision, retaining the activities that support good decisions and reliable delivery while reconsidering those that consume time without improving either.</p><p>Hermes gives me another way to conduct that examination. The interesting outcome isn’t how many agents I can keep busy. It’s whether the work they produce can be accepted with sufficient confidence, at a cost and pace that justify the arrangement.</p><h3>References</h3><p>[1] Beck, Kent, et al. <em>Manifesto for Agile Software Development</em>. 2001.<br><a href="https://agilemanifesto.org/">https://agilemanifesto.org/</a></p><p>[2] Nous Research. <em>Kanban — Multi-Agent Profile Collaboration</em>. Hermes Agent Documentation.<br><a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban">https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban</a></p><p>[3] Beck, Kent, et al. <em>Principles Behind the Agile Manifesto</em>. 2001.<br><a href="https://agilemanifesto.org/principles.html">https://agilemanifesto.org/principles.html</a></p><p>[4] Schwaber, Ken, and Jeff Sutherland. <em>The Scrum Guide: The Definitive Guide to Scrum</em>. November 2020.<br><a href="https://scrumguides.org/scrum-guide.html">https://scrumguides.org/scrum-guide.html</a></p><p>[5] Coleman, John, Daniel Vacanti, et al. <em>The Kanban Guide</em>. May 2025.<br><a href="https://kanbanguides.org/the-kanban-guide/">https://kanbanguides.org/the-kanban-guide/</a></p><p>[6] Toyota Motor Corporation. <em>Development and Deployment of the Toyota Production System</em>. In <em>75 Years of Toyota</em>.<br><a href="https://www.toyota-global.com/company/history_of_toyota/75years/text/entering_the_automotive_business/chapter1/section4/item4.html">https://www.toyota-global.com/company/history_of_toyota/75years/text/entering_the_automotive_business/chapter1/section4/item4.html</a></p><p>[7] Fowler, Martin. <em>The State of Agile Software in 2018</em>. Agile Australia keynote transcript. August 25, 2018.<br><a href="https://martinfowler.com/articles/agile-aus-2018.html">https://martinfowler.com/articles/agile-aus-2018.html</a></p><p>[8] Nous Research. <em>Kanban Tutorial</em>. Hermes Agent Documentation.<br><a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban-tutorial">https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban-tutorial</a></p><p>[9] Nous Research. <em>Kanban Worker Lanes</em>. Hermes Agent Documentation.<br><a href="https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban-worker-lanes">https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban-worker-lanes</a></p><p>[10] DORA and Google. <em>2025 State of AI-Assisted Software Development Report</em>. 2025.<br><a href="https://dora.dev/research/2025/dora-report/">https://dora.dev/research/2025/dora-report/</a></p><p>[11] DORA and Google Cloud. <em>2024 Accelerate State of DevOps Report</em>. 2024.<br><a href="https://dora.dev/research/2024/dora-report/">https://dora.dev/research/2024/dora-report/</a></p><p>[12] Jørgensen, Magne, and Eban Escott. “Relative Estimates of Software Development Effort: Are They More Accurate or Less Time-Consuming to Produce Than Absolute Estimates, and to What Extent Are They Person-Independent?” <em>Information and Software Technology</em>, vol. 143, 2022, article 106782.<br><a href="https://doi.org/10.1016/j.infsof.2021.106782">https://doi.org/10.1016/j.infsof.2021.106782</a></p><p>[13] Becker, Joel, Nate Rush, Beth Barnes, and David Rein. <em>Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity</em>. Model Evaluation &amp; Threat Research, July 10, 2025.<br><a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/</a></p><p>[14] Becker, Joel, Nate Rush, Tom Cunningham, David Rein, and Khalid Mahamud. <em>We Are Changing Our Developer Productivity Experiment Design</em>. Model Evaluation &amp; Threat Research, February 24, 2026.<br><a href="https://metr.org/blog/2026-02-24-uplift-update/">https://metr.org/blog/2026-02-24-uplift-update/</a></p><p>[15] Chen, Fiona, and James Stratton. <em>Artificial Intelligence in the Firm: Bottlenecks in Software Production</em>. Harvard University job market paper, August 4, 2026.<br><a href="https://fion.ac/jellyfish.pdf">https://fion.ac/jellyfish.pdf</a></p><p>[16] Anderson, David J. <em>A Brief History of Kanban for Knowledge Work</em>.<br><a href="https://djaa.com/brief-history-kanban-knowledge-work/">https://djaa.com/brief-history-kanban-knowledge-work/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d36513c90152" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Almost Learned Bazel in 2021. Do My iOS Apps Need It Now?]]></title>
            <link>https://medium.com/@raell.dottin/i-almost-learned-bazel-in-2021-do-my-ios-apps-need-it-now-f1abbe8b6deb?source=rss-fd9ac23f0a87------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1abbe8b6deb</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Raell Dottin]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 17:01:01 GMT</pubDate>
            <atom:updated>2026-11-07T17:01:01.879Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Five years after watching Bazel tutorials, I’m revisiting Google’s open-source build system through three SwiftUI applications of my own.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0ptdlgK2heFTu5WL" /><figcaption>Photo by <a href="https://unsplash.com/@hieu101193?utm_source=medium&amp;utm_medium=referral">Đào Hiếu</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Back in 2021, my employer was considering adopting Bazel as its build system. I remember watching engineers explain it in whiteboard videos circulating around Twitter. They would draw boxes, connect them with arrows, and explain how a build system could understand the relationships between components well enough to avoid repeating unnecessary work.</p><p>I found the approach interesting, although I never got far enough into Bazel to become comfortable using it. Watching someone explain a tool and remembering how to configure it several years later are different experiences. I understood the general idea, but without a project that required me to use Bazel regularly, most of what I learned eventually disappeared from memory.</p><p>Five years later, I have a better reason to revisit it.</p><p>I’m maintaining three applications written in Swift, Apple’s programming language: Gymphant, a workout planning and timing application; Owlory, a productivity application; and Dripnory, a weather-aware wardrobe planner. All three use SwiftUI, Apple’s framework for developing application interfaces, and each has accumulated its own collection of build scripts, automated tests, development conventions, and release procedures.</p><p>I’ve also become increasingly interested in using AI coding agents to help maintain these applications. An agent can investigate a bug, modify the code, execute tests, and use the results to guide another change. That workflow depends on the quality of the feedback it receives and how much time each validation cycle consumes.</p><p>Bazel seems relevant because it was designed to make software builds more predictable and avoid repeated work. Its dependency management and caching capabilities could be useful when automated development increases the frequency of compilation and testing [1][7].</p><p>However, adopting another build system introduces work of its own. I would need to maintain build definitions, verify compatibility with Apple’s development tools, and preserve the release processes that already work.</p><p>The decision depends on whether Bazel can eliminate enough repeated development work to justify those additional responsibilities. I can investigate that without replacing all three build systems at once.</p><h3>Understanding What Bazel Actually Does</h3><p>Every application needs a way to turn its source code into something people can run.</p><p>For Apple developers, Xcode handles most of that process. When I build an iPhone application, Xcode coordinates the Swift compiler, processes images and translations, links the application against Apple’s frameworks, and packages the resulting files into an application bundle. Depending on the application, it may also need to build an Apple Watch companion, widgets, or other extensions.</p><p>Much of this happens without requiring the developer to manage each individual operation. Xcode maintains information about which parts of the application depend on others and uses that information to organize the build.</p><p>Bazel approaches the same problem by making those relationships explicit.</p><p>Suppose I change a calculation in Gymphant that determines the duration of a rest period between workout sets. The calculation might be used by the iPhone application, the Apple Watch application, and several automated tests. Those components depend on the calculation, but an unrelated settings screen or collection of application icons has no reason to change.</p><p>A build system needs to determine which operations are affected by that modification.</p><p>Bazel represents these relationships through a <em>dependency graph</em>, a map showing how the components of a software project depend on one another. Each component that can be built or tested is represented by a named <em>target</em>. A target might be a Swift library, a collection of tests, or an entire iPhone application.</p><p>When source code changes, Bazel uses the dependency graph to identify the build operations that need to run. It can reuse previously generated results when the relevant inputs and build configuration have not changed [7].</p><p>This is called incremental building, and Xcode already supports it. Swift Package Manager, Apple’s tool for organizing Swift code into reusable packages, also performs incremental compilation.</p><p>Bazel’s potential advantage comes from the combination of explicit dependency relationships, fine-grained targets, and reusable build results. The benefits become more substantial when different builds repeatedly require the same components.</p><p>For example, imagine that three automated workflows independently compile the same unchanged library. If their build configurations and other relevant inputs match, a shared cache could allow later builds to reuse the earlier result rather than compiling the library again.</p><p>Bazel also emphasizes <em>hermetic builds</em>, which are designed to use explicitly declared inputs rather than depend on unexpected details of the machine performing the work. This reduces the likelihood that a build succeeds on one computer because of an undeclared local dependency and fails somewhere else. Apple’s compiler and software development tools still need to be configured consistently; Bazel does not make the platform requirements disappear.</p><p>The underlying engineering problem is understandable. I want changes to trigger the work necessary to validate them without repeatedly rebuilding components that have not changed.</p><p>Whether Bazel would perform that work more efficiently than my existing tools depends heavily on how the applications are organized.</p><h3>Three Applications, Three Different Build Problems</h3><p>I began by examining the repositories rather than assuming they would all benefit from the same migration.</p><p>That turned out to be useful because the three applications share a language and development platform, but their build requirements are quite different.</p><h3>Gymphant: An Existing Boundary for Experimentation</h3><p>Gymphant runs on iPhone and Apple Watch. It manages workout sessions, timers, progress tracking, and information shared between devices.</p><p>Its repository already contains a Swift package named GymphantCore. The package organizes a substantial amount of application logic separately from the complete iPhone interface and includes an exercise-library importer and automated tests [11].</p><p>This is relevant because I wouldn’t need to begin by translating the entire application into Bazel.</p><p>I could select a manageable part of the existing package, define equivalent Bazel targets, and compare compilation and testing against Swift Package Manager.</p><p>Gymphant also has an extensive Makefile, which provides convenient commands for building the application and running validation checks. Make is an older build-automation tool that executes commands according to declared targets and dependencies. In this repository, it coordinates Swift tests, iOS and watchOS builds, localization checks, performance validation, and release preparation [12].</p><p>Bazel would therefore be joining an environment that already has meaningful automation.</p><p>The experiment would need to establish whether Bazel provides better caching or more precise test execution, rather than merely providing another command for compiling the same code.</p><p>Gymphant’s existing package structure makes a limited experiment relatively approachable. Its Apple Watch integration and Live Activity extension would make a complete migration more involved.</p><h3>Owlory: A More Complicated Release Environment</h3><p>Owlory is a local-first productivity application that brings together daily planning, routines, reminders, writing, and progress tracking.</p><p>Its codebase is organized around domain rules, application services, persistence, and SwiftUI presentation. Like Gymphant, it has a reusable Swift package, OwloryCore, which allows selected application logic to be compiled and tested independently [13].</p><p>The complete product has considerably more platform integration.</p><p>Owlory includes Apple Watch components, widgets, Live Activities, and App Intents. App Intents is Apple’s framework for exposing application functionality to Siri, Shortcuts, and other system features. The application also supports numerous languages, adding translation resources and layout validation to its development process.</p><p>Its existing automation coordinates architectural checks, unit and interface tests, performance validation, localization, screenshot generation, and release preparation [14].</p><p>The build also needs to preserve the relationship between a release artifact and the source revision used to produce it. That relationship, commonly called <em>build provenance</em>, becomes important when investigating a problem in a particular version distributed through Apple’s TestFlight beta-testing service.</p><p>Bazel supports Apple application extensions and embedded targets, but successfully compiling Owlory’s Swift code would establish only part of the required compatibility [3].</p><p>I would also need to verify that the Watch application and widgets are packaged correctly, that their permissions remain intact, and that the release process produces the expected application identity.</p><p>Owlory has enough complexity to make better build management attractive. That same complexity increases the potential cost of migrating.</p><h3>Dripnory: A Simpler Application Packaging Experiment</h3><p>Dripnory is a weather-aware wardrobe application that combines forecasts, clothing management, outfit recommendations, and daily planning.</p><p>Its GitHub repository is named climate-closet-ios, although the application and Xcode scheme use the name Dripnory.</p><p>Compared with Gymphant and Owlory, its application structure is more straightforward. The repository contains an iPhone application, unit tests, integration tests, and user-interface tests, without the additional Watch application and widget targets found in the other projects [15].</p><p>These tests serve different purposes. Unit tests verify individual pieces of application logic. Integration tests check whether components work correctly together, while user-interface tests exercise the application through interactions resembling those of a person using it.</p><p>Dripnory’s Makefile already runs its test suites through Xcode and coordinates formatting checks, development tooling, and release validation [16].</p><p>The application also has separate development and release configurations, including different application identifiers and version metadata [17].</p><p>That makes Dripnory a reasonable candidate for testing whether Bazel can produce a complete iPhone application with the correct resources and configuration.</p><p>There is an architectural complication, though. Unlike Gymphant and Owlory, Dripnory does not currently expose its application logic through a separate Swift package. Adopting Bazel could require identifying new library boundaries or representing the existing application structure more directly.</p><p>I would want to avoid rearranging working application code solely to accommodate the build system.</p><p>These differences suggest two useful experiments. Gymphant could establish whether Bazel improves the development cycle for an existing library. Dripnory could establish whether Bazel can reproduce the behavior of a complete iPhone application.</p><p>Owlory would benefit from the lessons of both before I considered changing its development workflow.</p><h3>What It Would Take to Use Bazel</h3><p>The current Bazel tooling provides a reasonably clear route into Apple development.</p><p>Bazel itself coordinates the build. A companion tool called Bazelisk selects and launches the Bazel version specified by a project, which helps keep developer machines and automated build environments consistent [21].</p><p>Bazel gets its language-specific and platform-specific capabilities from collections of build rules. For my applications, the most relevant are rules_swift, which supports compiling Swift libraries and tests, and rules_apple, which handles Apple application packaging [3][4].</p><p>As of October 19, 2026, Bazel 9.3.0 is the stable version I would use for an initial evaluation [1]. The relevant Swift rules are available at version 4.2.1, while the Apple rules reached version 5.3.1 on October 17 [19][23].</p><p>That last update happened only three days after I began researching this article, when version 5.3.0 was current [18]. Build tooling changes frequently enough that I would want the repository to select known versions rather than silently use whatever happens to be installed.</p><p>A Bazel repository typically includes a file called MODULE.bazel, which declares external build dependencies using Bazel&#39;s module-management system, Bzlmod.</p><p>An illustrative configuration might contain:</p><pre># MODULE.bazel — illustrative configurationmodule(name = &quot;gymphant_bazel_trial&quot;)bazel_dep(name = &quot;rules_swift&quot;, version = &quot;4.2.1&quot;)bazel_dep(name = &quot;rules_apple&quot;, version = &quot;5.3.1&quot;)</pre><p>This declares which build-rule packages are required. It does not compile Gymphant or configure its application targets.</p><p>Those targets are described in files conventionally named BUILD or BUILD.bazel, using Starlark, a language with syntax resembling Python.</p><p>For example, an isolated Swift library could be described as follows:</p><pre># BUILD.bazel — illustrative targetload(&quot;@rules_swift//swift:swift.bzl&quot;, &quot;swift_library&quot;)swift_library(    name = &quot;WorkoutRules&quot;,    srcs = [&quot;Examples/WorkoutRules.swift&quot;],)</pre><p>The example assumes that Examples/WorkoutRules.swift exists and contains code that can compile as an independent library. Neither that file nor the target represents a completed change to Gymphant.</p><p>Once configured, the target could be compiled with:</p><pre>bazel build //:WorkoutRules</pre><p>The expression //:WorkoutRules identifies the target named WorkoutRules in the repository&#39;s root directory. Other targets could depend on it, and a separate test target could verify its behavior. The official Bazel Apple tutorial uses this general approach to explain how Swift libraries become dependencies of complete applications [2].</p><p>There is also tooling for consuming external Swift packages through Bazel, which could be useful if my applications acquire more third-party dependencies [8].</p><p>The appeal is that build relationships become explicit and available to Bazel’s dependency analysis. However, creating the rules is only the beginning. A complete application still needs configuration for resources, framework dependencies, application identifiers, signing, and any embedded extensions.</p><p>Developers can also use rules_xcodeproj to generate an Xcode project from Bazel targets. Version 4.1.0 of that integration was released in May 2026 [5][20].</p><p>That would allow me to retain familiar Xcode development workflows while experimenting with Bazel as the underlying build system, provided the necessary Xcode features work correctly.</p><p>The examples above are explanations of how Bazel is configured, not instructions that have been tested against my repositories. A working migration would require examining the source dependencies and existing Xcode settings before defining production targets.</p><h3>The Case for Adopting Bazel Earlier</h3><p>There is a reasonable argument for introducing Bazel before any of these applications becomes substantially larger.</p><p>A build system influences how developers organize code. If Gymphant, Owlory, and Dripnory eventually adopt consistent build conventions, sharing validation workflows across them could become easier. Bazel’s explicit dependency definitions could also encourage clearer boundaries between application logic and platform-specific code.</p><p>Starting early might avoid migrating a much larger collection of targets later.</p><p>This argument becomes more interesting when considering the automation I already maintain. Although the applications are separate products, I increasingly want common approaches to testing, release verification, and agent-assisted development.</p><p>A standardized build interface could make those workflows easier to coordinate.</p><p>There are limits to that benefit.</p><p>The applications already expose much of their validation through Makefiles. A shared automation tool can invoke make test or another established command without requiring every repository to use the same underlying build system.</p><p>Gymphant and Owlory also have reusable Swift packages, so some of the architectural benefits associated with smaller, independently testable components are already available.</p><p>Standardizing on Bazel would therefore need to provide something beyond consistent command names. Better cache reuse, more reliable dependency tracking, or substantially cheaper repeated validation would be stronger reasons.</p><p>The cost of waiting is also uncertain. Migrating a more complicated application could become harder later, but introducing a second build configuration today creates maintenance work immediately.</p><p>I would need evidence that the future migration cost is likely to exceed the cost of maintaining Bazel throughout the intervening development.</p><p>That is difficult to establish without understanding how much build work is currently being repeated.</p><h3>What Slack’s Bazel Migration Tells Me About Scale</h3><p>Slack provides a useful case study because its mobile engineering organization has documented the problems that led it to adopt Bazel.</p><p>In <em>Scaling Slack’s Mobile Codebases: Modularization</em>, Slack described reorganizing large mobile applications into smaller modules. Its engineers had encountered increasing complexity as more code accumulated in shared application targets.</p><p>The existing Xcode project files also created coordination problems. Adding frameworks could involve complicated project changes, and developers frequently encountered conflicts when multiple people modified those files.</p><p>Bazel allowed Slack to describe targets and dependencies separately from the generated Xcode project. The company also invested in tooling to make creating and maintaining modules easier [9].</p><p>That addressed an organizational problem I do not currently experience at the same scale.</p><p>I’m not coordinating dozens of engineers who regularly modify a shared Xcode project. Avoiding project-file conflicts is useful, but it would not be a primary reason for me to adopt Bazel.</p><p>Slack’s experience with build caching is more relevant.</p><p>The company found that dividing an application into smaller modules allowed Bazel to reuse more build results. Its engineers also recognized that caching provides relatively little benefit when most application code remains inside one large target that must be rebuilt after frequent changes [9].</p><p>That observation applies directly to my applications.</p><p>Gymphant and Owlory already have independently testable Swift packages, which could provide useful boundaries for Bazel. However, those same packages also allow me to run selected tests without Bazel. Their existence creates an opportunity for comparison without establishing which build system would perform better.</p><p>Slack later described using Bazel’s cache and bazel-diff, a tool that identifies which build targets are affected by changes between Git revisions. That allowed its continuous integration system to concentrate testing on affected components rather than repeatedly executing every test.</p><p>Continuous integration is the automated process of building and testing software changes, typically before accepting them into a project’s main development branch.</p><p>In its 2022 engineering account, Slack reported reducing its average time to merge to approximately nine minutes, with some changes completing the process in four and a half minutes [10].</p><p>Those results reflect a larger engineering effort involving modularization, build infrastructure, and selective testing. They cannot be attributed to Bazel alone, and they do not predict what would happen in my repositories.</p><p>What I take from Slack’s experience is that the structure of the software and the frequency of repeated work influence the value of the build system.</p><p>An organization with many developers may benefit from eliminating repeated compilation across hundreds of changes. An independent developer could also benefit if automated workflows repeatedly build the same components. Conversely, a substantial application with an efficient existing build process might gain relatively little.</p><p>Developer headcount is only one part of the calculation.</p><h3>AI Coding Agents Change the Workload, but Not Automatically the Answer</h3><p>This is where my current development workflow introduces a consideration I did not have in 2021.</p><p>I use AI coding agents to help investigate problems, implement changes, and verify application behavior. An agent might modify workout scheduling logic in Gymphant, execute the associated tests, examine a failure, and adjust its implementation.</p><p>Another task might involve Owlory’s reminder scheduling or Dripnory’s outfit recommendations.</p><p>These workflows can involve repeated compilation and testing, particularly when several attempts are needed to produce an acceptable change.</p><p>The relevant cost is the amount of work performed during each iteration.</p><p>Suppose an agent modifies one part of Gymphant and runs the complete application test suite after every change, even when a smaller collection of tests would provide sufficient feedback.</p><p>Bazel might make it easier to identify affected components and reuse cached outputs. However, I could also improve that workflow by teaching the agent to run the appropriate existing Swift Package Manager tests.</p><p>The latter might require substantially less engineering effort.</p><p>There is another complication: caching is most useful when the required inputs remain unchanged. If an agent repeatedly modifies the same library, the compiler may need to rebuild that library regardless of which build system is responsible.</p><p>Separate agents working in isolated checkouts may also have different build configurations, source revisions, or compiler settings. Their outputs would not necessarily be interchangeable.</p><p>A shared cache could still help when the inputs match, but the number of agents does not determine how often that happens.</p><p>Bazel’s remote caching could eventually allow compatible development environments to share compiled outputs [7]. That introduces additional infrastructure, including a cache service, authentication, storage management, and security considerations. Cached artifacts can contain application binaries or other sensitive build outputs, so access to that infrastructure would need to be controlled.</p><p>For now, I would evaluate local caching before adding a remote service.</p><p>The possibility I find interesting is a development environment in which agents can validate changes quickly because the build system understands which components require work.</p><p>I would still need to measure how much unnecessary work they are actually performing.</p><h3>The Apple Tooling Is Still a Reason for Caution</h3><p>Even if Bazel improves library compilation, I would need to establish that it works with the Apple development features my applications depend on.</p><p>Gymphant needs its Apple Watch companion and Live Activity extension packaged correctly. Owlory has additional requirements involving widgets, App Intents, localization, and application capabilities. Dripnory must preserve its asset catalogs, translated strings, application identifiers, and separate development and release configurations.</p><p>An <em>application extension</em> is a separately packaged component that allows an application to provide functionality through features such as widgets. These extensions need to be built, configured, and embedded correctly for the operating system to recognize them.</p><p>The official Apple rules support many of these requirements [3]. However, the supported build-rule features and the integration with Xcode are separate compatibility concerns.</p><p>The rules_swift documentation describes support for contemporary Xcode releases, including Xcode 26. Meanwhile, the published rules_xcodeproj compatibility matrix lists officially tested Xcode versions through 16.x [4][5].</p><p>That does not establish that Xcode 26 is incompatible with generated projects. It means the published compatibility information does not provide sufficient assurance for the toolchain I use.</p><p>SwiftUI previews introduce another concern.</p><p>Previews allow developers to examine interface changes inside Xcode without running the complete application. They are useful when adjusting layouts, typography, and interface components.</p><p>Although rules_xcodeproj lists previews among its features, an open issue documents problems with native SwiftUI previews. A related pull request proposes dedicated configurations intended to address those limitations [6][22].</p><p>I would need to test previews, code navigation, debugging, and simulator execution before relying on Bazel-generated Xcode projects for daily development.</p><p>The release process presents a separate set of requirements.</p><p>My existing workflows verify application versions, signing, and build provenance. Code signing attaches a digital signature that allows Apple platforms to verify the identity and integrity of an application.</p><p>A successful simulator build would demonstrate that Bazel can compile and package an application for that environment. It would not establish that the same configuration can produce a properly signed application for TestFlight or the App Store.</p><p>I would therefore keep the existing Xcode-based release workflows authoritative during the initial experiments. The Bazel implementation would have to demonstrate equivalent behavior before assuming responsibility for distributing an application.</p><p>This approach also gives me a way to investigate compatibility without placing working releases at risk.</p><h3>How I Would Decide Whether Bazel Is Worth Maintaining</h3><p>The central question is how much useful development time Bazel could recover.</p><p>I can express the recurring benefit in reasonably simple terms:</p><p><strong>Recurring time saved = number of eligible build operations × average time avoided per operation.</strong></p><p>The word <em>eligible</em> is important. Bazel can only reuse a cached result when the relevant inputs and configuration match. The calculation therefore needs to account for how often cache reuse actually occurs, rather than assuming every build becomes faster.</p><p>The savings must then be compared with the additional maintenance work.</p><p>That includes updating build definitions, accommodating new Apple platform features, maintaining compatibility with Xcode releases, and investigating failures specific to Bazel.</p><p>There is also an initial migration cost. It may be substantial even if the eventual build becomes faster.</p><p>I would consider three questions when making the decision.</p><p>First, how often do I repeat the work Bazel could eliminate? A small improvement in an operation performed hundreds of times may be more useful than a large improvement in a build performed once a week.</p><p>Second, how much of the application can be compiled and tested independently? A codebase divided into stable modules creates more opportunities to reuse unaffected components. An application that frequently rebuilds one large target may benefit less.</p><p>Third, what would I have to maintain to preserve those savings? If adopting Bazel requires frequent adjustments to platform rules, application packaging, and generated Xcode projects, the recurring maintenance cost could offset the faster builds.</p><p>Not all costs can be reduced to elapsed time. A missing entitlement or incorrectly packaged Watch application could create a release failure whose consequences are much greater than the compilation time saved.</p><p>Release correctness therefore needs to be a requirement that any proposed migration satisfies, rather than another number averaged into a performance comparison.</p><p>This framework also explains why I would not adopt Bazel merely to standardize the three repositories.</p><p>The applications are independently maintained and do not currently share one large common codebase. Bazel would not automatically allow them to reuse compiled outputs simply because they use the same programming language. Cache reuse depends on equivalent build operations and inputs, not the names of the projects.</p><p>Common build conventions could still be valuable, but my existing automation already provides a way to invoke different development tools through consistent commands.</p><p>I would want the improvements to come from work genuinely avoided.</p><h3>The Experiments I Would Actually Run</h3><p>I haven’t yet completed a Bazel migration for Gymphant, Owlory, or Dripnory. The conclusions here come from examining the existing repositories, studying the tooling, and comparing the problems Bazel addresses with the work my applications currently perform.</p><p>The next step would be to test two separate hypotheses.</p><p><strong>The first is that Bazel can improve the feedback cycle for an existing Swift library.</strong></p><p>Gymphant is the appropriate starting point because GymphantCore already provides a boundary between reusable application logic and the complete Apple application.</p><p>I would select a small part of its workout domain logic and identify the tests that verify its behavior.</p><p>Before introducing Bazel, I would record the existing compilation and test times using Swift Package Manager.</p><p>I would then define equivalent Bazel targets and verify that both systems compile the selected code and execute the same meaningful tests.</p><p>Only after establishing correctness would I measure performance.</p><p>I would compare clean builds, incremental builds after a small Swift change, repeated builds with available cached outputs, and test execution times. The measurements would need to use comparable environments and be repeated enough to distinguish consistent improvements from ordinary variation.</p><p>I would also record cache hits and misses. A faster build does not necessarily mean caching caused the improvement, particularly when the two build systems organize compilation differently.</p><p>If Bazel fails to produce a measurable improvement over the existing Swift package workflow, that would be a reason to stop the experiment rather than proceed directly to a more complicated migration.</p><p><strong>The second hypothesis is that Bazel can reproduce a complete application’s behavior without disrupting the development workflow.</strong></p><p>Dripnory would be the candidate for that experiment.</p><p>I would configure Bazel to produce an iOS Simulator application, then compare its behavior with the existing Xcode build. The application would need to launch correctly, load its assets and translations, preserve the expected development identity, and pass its established unit, integration, and interface tests.</p><p>The tests would need to verify equivalent behavior rather than simply demonstrate that both build systems return successful exit codes.</p><p>I would also test the normal debugging workflow. If the Bazel-generated application builds successfully but becomes inconvenient to inspect or debug in Xcode, that would affect whether I wanted to use it regularly.</p><p>The initial objective would be compatibility. Performance comparisons would follow only after that requirement was satisfied.</p><p>Gymphant’s complete iPhone and Apple Watch integration could then be considered, using the lessons from Dripnory’s packaging experiment.</p><p>Owlory would come later because its broader collection of platform features increases the work required to establish equivalent behavior.</p><p>At each stage, I would compare Bazel against improvements available through the existing build system. Better-targeted tests, clearer module boundaries, and more selective automated validation may deliver some of the same benefits without introducing another configuration to maintain.</p><p>A successful experiment would justify further investigation, but it would not automatically justify migrating all three applications.</p><h3>A Better Reason to Learn Bazel</h3><p>When I first encountered Bazel in 2021, I was interested in the technology without having a project that required me to understand it.</p><p>I could follow the whiteboard explanations, but I never developed the familiarity that comes from using a tool to solve an actual problem.</p><p>Five years later, I have three applications with established build configurations, automated tests, Apple-platform integrations, and release processes. Their development also involves AI coding agents that repeatedly depend on those systems for reliable feedback.</p><p>I now have a practical way to investigate what Bazel could contribute.</p><p>The research has also changed how I think about the decision. Application size alone is a poor reason to adopt a more elaborate build system. The stronger considerations are how often development workflows repeat work, how much of that work can be reused, and what the additional infrastructure costs to maintain.</p><p>Bazel may be valuable in that environment, particularly if automated development continues increasing the number of build-and-test cycles I perform. The existing Swift packages also give me a relatively low-risk way to investigate its capabilities.</p><p>For now, I’m keeping the build systems that already work.</p><p>I’ll start with a small part of Gymphant, measure the results, and expand the experiment only if the evidence supports it. If Bazel proves useful, I’ll have a reason to maintain it. If the existing tools perform just as well, I’ll have avoided replacing reliable infrastructure without a clear benefit.</p><p>Either result would teach me more than another afternoon of watching whiteboard videos. This time, I’ll have something I didn’t have in 2021: a way to determine whether Bazel solves a problem I actually have.</p><h3>References</h3><p>[1] Bazel Project. <em>Bazel Releases</em>. Official release history, including Bazel 9.3.0, released October 7, 2026.<br><a href="https://github.com/bazelbuild/bazel/releases?utm_source=chatgpt.com">https://github.com/bazelbuild/bazel/releases</a></p><p>[2] Bazel Apple Rules Maintainers. <em>Bazel Tutorial: Build an iOS App</em>. Official tutorial covering Swift libraries, build targets, iOS application packaging, and Xcode integration.<br><a href="https://github.com/bazelbuild/rules_apple/blob/main/doc/tutorials/ios-app.md?utm_source=chatgpt.com">https://github.com/bazelbuild/rules_apple/blob/main/doc/tutorials/ios-app.md</a></p><p>[3] Bazel Apple Rules Maintainers. <em>Bazel Rules for iOS Applications and Bundles</em>. Official reference documentation for iOS applications, extensions, resources, and packaging.<br><a href="https://github.com/bazelbuild/rules_apple/blob/main/doc/rules-ios.md?utm_source=chatgpt.com">https://github.com/bazelbuild/rules_apple/blob/main/doc/rules-ios.md</a></p><p>[4] Bazel Swift Rules Maintainers. <em>Swift Rules for Bazel</em>. Official documentation for Swift compilation, testing, and toolchain compatibility.<br><a href="https://github.com/bazelbuild/rules_swift?utm_source=chatgpt.com">https://github.com/bazelbuild/rules_swift</a></p><p>[5] Mobile Native Foundation. <em>rules_xcodeproj: Bazel Rules for Generating Xcode Projects</em>. Documentation covering Xcode integration, supported features, and tested compatibility.<br><a href="https://github.com/MobileNativeFoundation/rules_xcodeproj?utm_source=chatgpt.com">https://github.com/MobileNativeFoundation/rules_xcodeproj</a></p><p>[6] Mobile Native Foundation. <em>Feature Request: Xcode SwiftUI Preview Support</em>. GitHub Issue #3201, opened June 20, 2025.<br><a href="https://github.com/MobileNativeFoundation/rules_xcodeproj/issues/3201?utm_source=chatgpt.com">https://github.com/MobileNativeFoundation/rules_xcodeproj/issues/3201</a></p><p>[7] Bazel Project. <em>Remote Caching</em>. Official documentation explaining cached build outputs, cache hits, and remote caching infrastructure.<br><a href="https://bazel.build/remote/caching?utm_source=chatgpt.com">https://bazel.build/remote/caching</a></p><p>[8] Bazel Central Registry. <em>rules_swift_package_manager</em>. Module documentation for integrating external Swift packages into Bazel builds.<br><a href="https://registry.bazel.build/modules/rules_swift_package_manager/?utm_source=chatgpt.com">https://registry.bazel.build/modules/rules_swift_package_manager/</a></p><p>[9] Slack Engineering. <em>Scaling Slack’s Mobile Codebases: Modularization</em>. Engineering case study covering modularization, Bazel adoption, dependency management, and development tooling.<br><a href="https://slack.engineering/stabilize-modularize-modernize-scaling-slacks-mobile-codebases-2/?utm_source=chatgpt.com">https://slack.engineering/stabilize-modularize-modernize-scaling-slacks-mobile-codebases-2/</a></p><p>[10] Slack Engineering. <em>Mobile Developer Experience at Slack</em>. October 12, 2022; updated October 19, 2022. Engineering case study discussing Bazel caching, selective testing, and continuous integration performance.<br><a href="https://slack.engineering/mobile-developer-experience-at-slack/?utm_source=chatgpt.com">https://slack.engineering/mobile-developer-experience-at-slack/</a></p><p>[11] Dottin, Raell. <em>Gymphant — Package.swift</em>. Swift Package Manager configuration, application library boundaries, and test targets. Private GitHub repository.<br><a href="https://github.com/raelldottin/gymphant/blob/main/Package.swift">https://github.com/raelldottin/gymphant/blob/main/Package.swift</a></p><p>[12] Dottin, Raell. <em>Gymphant — Makefile</em>. Existing build, testing, validation, and release commands. Private GitHub repository.<br><a href="https://github.com/raelldottin/gymphant/blob/main/Makefile">https://github.com/raelldottin/gymphant/blob/main/Makefile</a></p><p>[13] Dottin, Raell. <em>Owlory — Package.swift</em>. OwloryCore package configuration and associated test targets. Private GitHub repository.<br><a href="https://github.com/raelldottin/owlory/blob/main/owlory_xcode/Package.swift">https://github.com/raelldottin/owlory/blob/main/owlory_xcode/Package.swift</a></p><p>[14] Dottin, Raell. <em>Owlory — Makefile</em>. Application validation, localization, performance, testing, and release workflows. Private GitHub repository.<br><a href="https://github.com/raelldottin/owlory/blob/main/Makefile">https://github.com/raelldottin/owlory/blob/main/Makefile</a></p><p>[15] Dottin, Raell. <em>Dripnory — README.md</em>. Application architecture, development requirements, testing strategy, and release configuration. Private GitHub repository.<br><a href="https://github.com/raelldottin/climate-closet-ios/blob/main/README.md">https://github.com/raelldottin/climate-closet-ios/blob/main/README.md</a></p><p>[16] Dottin, Raell. <em>Dripnory — Makefile</em>. Existing Xcode build commands, automated tests, and development checks. Private GitHub repository.<br><a href="https://github.com/raelldottin/climate-closet-ios/blob/main/Makefile">https://github.com/raelldottin/climate-closet-ios/blob/main/Makefile</a></p><p>[17] Dottin, Raell. <em>Dripnory — Config/App.xcconfig</em>. Application identifiers, build configurations, and version metadata. Private GitHub repository.<br><a href="https://github.com/raelldottin/climate-closet-ios/blob/main/Config/App.xcconfig">https://github.com/raelldottin/climate-closet-ios/blob/main/Config/App.xcconfig</a></p><p>[18] Bazel Central Registry. <em>rules_apple 5.3.0</em>. Published October 9, 2026. Apple-platform build rules examined during the initial research.<br><a href="https://registry.bazel.build/modules/rules_apple/5.3.0/?utm_source=chatgpt.com">https://registry.bazel.build/modules/rules_apple/5.3.0/</a></p><p>[19] Bazel Central Registry. <em>rules_swift 4.2.1</em>. Published October 9, 2026. Swift compilation and testing rules.<br><a href="https://registry.bazel.build/modules/rules_swift/4.2.1/?utm_source=chatgpt.com">https://registry.bazel.build/modules/rules_swift/4.2.1/</a></p><p>[20] Mobile Native Foundation. <em>rules_xcodeproj 4.1.0</em>. Released May 7, 2026. Xcode project generation rules for Bazel.<br><a href="https://github.com/MobileNativeFoundation/rules_xcodeproj/releases/tag/4.1.0?utm_source=chatgpt.com">https://github.com/MobileNativeFoundation/rules_xcodeproj/releases/tag/4.1.0</a></p><p>[21] Bazel Project. <em>Bazelisk</em>. Official project documentation covering Bazel version selection and developer setup.<br><a href="https://github.com/bazelbuild/bazelisk?utm_source=chatgpt.com">https://github.com/bazelbuild/bazelisk</a></p><p>[22] Mobile Native Foundation. <em>Support Native Xcode Previews in Dedicated Configurations</em>. GitHub Pull Request #3325, opened August 6, 2026. Development work addressing native SwiftUI preview integration.<br><a href="https://github.com/MobileNativeFoundation/rules_xcodeproj/pull/3325">https://github.com/MobileNativeFoundation/rules_xcodeproj/pull/3325</a></p><p>[23] Bazel Central Registry. <em>rules_apple 5.3.1</em>. Published October 17, 2026. Updated Apple-platform build rules.<br><a href="https://registry.bazel.build/modules/rules_apple/5.3.1/?utm_source=chatgpt.com">https://registry.bazel.build/modules/rules_apple/5.3.1/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1abbe8b6deb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>