<?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 Oliver Tappin on Medium]]></title>
        <description><![CDATA[Stories by Oliver Tappin on Medium]]></description>
        <link>https://medium.com/@olivertappin?source=rss-5eca2b78a00b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*nOrtqtYciIliTeXMWAIjlw.jpeg</url>
            <title>Stories by Oliver Tappin on Medium</title>
            <link>https://medium.com/@olivertappin?source=rss-5eca2b78a00b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 07:16:35 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@olivertappin/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[Stop Using AI for Problems Automation Already Solved]]></title>
            <link>https://olivertappin.medium.com/stop-using-ai-for-problems-automation-already-solved-a1698e632a0b?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1698e632a0b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Thu, 11 Jun 2026 11:01:41 GMT</pubDate>
            <atom:updated>2026-06-11T11:01:41.880Z</atom:updated>
            <content:encoded><![CDATA[<h3>As AI pricing shifts from seats to consumption, the companies that win won’t be the ones using AI everywhere. They’ll be the ones that know when not to.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jtSeoKTk-5_BL-WN" /><figcaption>Photo by <a href="https://unsplash.com/@ibrahimboran?utm_source=medium&amp;utm_medium=referral">Ibrahim Boran</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>There is a strange pressure in the market at the moment. It is no longer enough to say that your company uses AI, or that your teams are experimenting with AI, or that you have a handful of sensible AI use cases in progress. The new expectation is that AI should be everywhere. It should be in every workflow, every product, every team, every customer journey, every back-office process and every board update. If a piece of software does not have an AI assistant, it starts to feel unfinished. If a strategy deck does not mention agents, copilots or automation, it can feel out of date before it has even been presented.</p><p>That pressure is understandable. AI adoption has moved incredibly quickly. <a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai">McKinsey’s 2025 State of AI survey</a> found that 88% of respondents said their organisations were using AI regularly in at least one business function, up from 78% the year before. At the same time, only around a third said their organisations had begun to scale AI across the enterprise, and just 39% reported enterprise-level EBIT impact. In other words, AI is being used widely, but not always deeply, and not always in ways that translate into measurable business value.</p><p>That gap really matters, because the most important question for leaders is no longer whether AI is powerful. It clearly is. The more useful question is whether we are applying it with enough discipline.</p><p>Nearly six years ago, I wrote an article called <a href="https://olivertappin.medium.com/4-reasons-why-you-need-automation-e54d2f8061ca"><em>4 Reasons Why You Need Automation</em></a>. The argument was simple: repeated manual work creates avoidable errors, wastes time, hides knowledge in people’s heads and stops teams from focusing on more valuable work. The examples and tooling around that argument have aged, but the principle has not. If anything, the arrival of generative AI has made the principle more important rather than less.</p><p>What concerns me about the current AI conversation is not that people are excited. They should be. What worries me is that in the rush to use AI everywhere, we seem to have forgotten what automation was for in the first place. We learned, over many years, that the right way to improve a process was to understand it, simplify it, remove unnecessary human intervention, make it repeatable, measure it and only then decide where technology could help. Now, too many conversations start one step later. The question becomes “How do we add AI to this?” rather than “What is the problem we are actually trying to solve?”</p><p>That may sound like a subtle difference, but in operations it is everything. A badly understood process does not become a good process because a language model has been added to it. A confusing workflow does not become elegant because it has an AI assistant on the front. A poor decision tree does not become intelligent because the final response is written in natural language. AI can make a good system more useful, but it can also make a bad system more expensive, harder to understand and more difficult to govern.</p><p>One of the most interesting examples of this is the rise of AI “skills”. Anthropic describes Claude Skills as reusable resources that provide workflows, context and best practices, allowing Claude to perform specialised tasks without users having to repeat the same guidance in every conversation. Claude’s own documentation gives examples such as generating emails, applying brand guidelines and integrating with tools like Jira or Linear.</p><p>That is a genuinely useful idea. It is also, in a way, a return to something very familiar. We are packaging instructions. We are standardising inputs. We are providing scripts, templates and reference material. We are trying to make an inherently flexible system produce consistent outcomes. The language around it is new, but the operational instinct is old: when something needs to happen more than once, give it structure.</p><p>I do not see that as a criticism of AI. I see it as evidence of maturity. The more powerful AI becomes, the more we need the boring disciplines around it. Without those disciplines, every interaction becomes a one-off. Every result depends on the phrasing of a prompt, the state of the context window, the model chosen, the documents attached and the mood of the user. That may be acceptable for exploration, but it is a weak foundation for a business process.</p><p>This is where automation still earns its place. Automation is not unfashionable simply because AI is more exciting. In many cases, automation is exactly what makes AI useful. It provides the rails, the boundaries and the repeatability. It decides when something should happen, what data should be passed in, what checks should be applied, what output is acceptable and what should happen next. AI may perform the reasoning or language-heavy part of the task, but automation often makes the task safe enough and consistent enough to run in the real world.</p><p>The distinction becomes clearer when you think about the most dependable technologies around us. A heating timer does not need to understand your lifestyle aspirations. A thermostat does not need to hold a conversation about comfort. Traffic lights do not need to produce a beautifully reasoned essay about vehicle flow at every junction. These systems are valuable because they are predictable. They are narrow, constrained and extremely good at doing the same thing again and again.</p><p>Of course, modern versions of these systems can be made smarter. A smart thermostat can learn behaviour, optimise energy usage and respond to changing conditions. Traffic management systems can use sensors, modelling and adaptive control to improve flow. There are legitimate cases where intelligence adds value. But the important point is that intelligence is not the starting point. The starting point is the job to be done. If the job is to switch something on at a fixed time, a schedule is probably enough. If the job is to maintain a target temperature, a simple control loop may be enough. If the job is to remind someone every day at midday to submit a report, an AI agent is almost certainly overkill.</p><p>That is the test we are not applying often enough. Does this task require reasoning, judgement, language understanding or adaptation? Or does it require consistency, timing, validation and execution? Those are different categories of work. The first may benefit from AI. The second is usually better served by automation.</p><p>Forrester made a related point in its 2025 automation predictions, arguing that generative AI would orchestrate less than 1% of core business processes that year, while deterministic automation would continue to control the core, long-running process and AI would contribute bursts of insight or efficiency around it. Its 2026 automation outlook is similarly cautious, suggesting that ROI and governance challenges will keep most organisations running deterministic automation even as vendors push agentic features.</p><p>That feels much closer to what mature AI adoption should look like. Not AI everywhere as a blanket rule, but AI placed carefully where ambiguity, judgement and language create genuine friction. Not agents roaming freely across every operational workflow, but well-designed systems where the deterministic parts remain deterministic and the probabilistic parts are used deliberately.</p><p>The commercial side of this is becoming just as important as the technical side. For the first phase of enterprise AI adoption, the pricing model often felt familiar: buy seats, give people access, encourage usage, measure adoption. That model makes sense when the goal is to get people experimenting. It is also psychologically powerful. Once users build habits around a tool, the organisation becomes more dependent on it. The more work that moves into the tool, the harder it becomes to unwind later.</p><p>But AI economics do not behave like traditional software economics. A normal SaaS product can often serve another user at relatively low marginal cost. AI is different. Every prompt, every response, every document passed into context, every image generated, every reasoning step and every agent action consumes compute. That does not mean AI is bad value. It does mean that the cost model is more like metered infrastructure than conventional office software.</p><p>You can already see the market moving in that direction. OpenAI’s own documentation now describes flexible pricing for ChatGPT Business, Enterprise and Edu, where credits unlock additional access to advanced features such as Deep Research, Thinking models, Image Generation, Advanced Voice and Codex. Enterprise and Edu workspaces can purchase a shared credit pool at the contract level, with users and seat types drawing from that pool for advanced features.</p><p>OpenAI has also introduced pay-as-you-go pricing for Codex seats in ChatGPT Business and Enterprise, with usage billed on token consumption rather than through a fixed seat fee. The stated benefit is that teams can see more clearly how usage turns into spend across budgets, workflows and teams.</p><p>This is not an isolated OpenAI story. McKinsey’s work on AI and SaaS business models argues that while traditional per-user subscriptions are unlikely to disappear, AI capabilities will increasingly require some form of consumption-based pricing in the mix. The reason is straightforward: as software starts performing work rather than merely supporting work, value is less correlated with the number of humans logging in and more correlated with the volume and complexity of work completed.</p><p>Salesforce now advertises Agentforce pricing that includes consumption-based options through Flex Credits or Conversations, alongside per-user licensing. Zendesk has pushed further into outcome-based AI pricing, with AI agent pricing tied to automated resolutions. Intercom’s Fin AI Agent pricing is also outcome-based, charging per outcome rather than simply per seat. These examples differ in detail, but they all point in the same direction: the industry is searching for pricing models that connect AI cost to usage, work or value rather than just access.</p><p>Between drafting this article and publishing it, the most telling example yet arrived. On 1 June 2026, GitHub <a href="https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/">moved every Copilot plan to usage-based billing</a>. The headline subscription prices did not change, but the familiar allowance of premium requests was replaced by GitHub AI Credits, metered against the tokens each interaction consumes and priced at the published rates of whichever model handles the task. Code completions remain unlimited; chat, agents and everything else now draw down a monthly allotment, with pooled credits and budget controls for business customers. GitHub’s reasoning was candid enough: agentic workflows consume far more compute than flat-rate pricing was ever built to absorb.</p><p>The reaction was immediate. Within hours of the switchover, developers were reporting that perfectly ordinary working sessions were consuming startling portions of a month’s allowance. One subscriber on the $39 Pro+ plan calculated that <a href="https://www.theregister.com/ai-and-ml/2026/06/02/github-copilot-users-threaten-exit-as-metered-billing-kicks-in/5249826">two hours of routine development work had used 8%</a> of their monthly credits. Others watched a full month’s allotment disappear in a day, and projections of bills rising <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/github-copilot-customers-suffer-from-sticker-shock-as-microsoft-switches-to-usage-based-pricing-customers-report-up-to-100-fold-price-hikes">tenfold, fiftyfold, in some reports a hundredfold</a>, spread across <a href="https://github.com/orgs/community/discussions/192948">GitHub’s community forum</a>, Reddit and X. The detail I find most telling is the reason why. Long-running chat threads <a href="https://www.techspot.com/news/112628-github-switched-copilot-metered-billing-developers-watching-months.html">resend their entire history as input tokens</a> with every message. Agents fed sprawling context consume credits at a rate nobody previously had to notice. The meter did not create those costs. It revealed them.</p><p>For executives and operations leaders, this should change the conversation. A per-seat licence encourages adoption. A consumption model demands discipline. Once usage is metered, the question is no longer “How many people have access?” It becomes “What are they using it for, how often, with what inputs, at what cost, and with what measurable result?”</p><p>That is where some organisations may get a shock. Copilot’s users have just had theirs. The early habit many people develop with AI is to give the model everything. More documents. More background. More instructions. More historical examples. More meeting notes. More policies. More prompts layered on top of prompts. When the answer is weak, the instinct is often to add even more context.</p><p>Sometimes that is exactly what is needed. Often it is not. More context can improve an answer, but it can also make the system slower, more expensive and harder to reason about. It can hide the real task inside a pile of loosely related information. It can turn a simple workflow into a sprawling conversation. It can also create a new kind of operational debt: not code debt, but context debt.</p><p>Context debt appears when nobody can explain why an AI workflow behaves the way it does because the behaviour is buried in accumulated prompts, files, examples, memories and exceptions. It is what happens when a team keeps feeding a model more background instead of redesigning the underlying process. It is what happens when AI becomes the place where complexity goes to hide. Under metered billing, context debt stops being an abstraction. It becomes a line on the invoice.</p><p>The antidote is not to avoid AI. The antidote is to decompose the work properly. Before asking a model to do something, break the process down into its smallest useful parts. Identify which steps are deterministic and should be automated. Identify which steps require human approval. Identify which steps require judgement or interpretation. Identify what data is truly necessary and what is merely noise. Then, and only then, decide where AI belongs.</p><p>That kind of decomposition may feel slower at first, but it is exactly what prevents runaway cost and operational fragility later. It also produces better AI use cases. Instead of asking a model to “handle customer onboarding”, you might ask it to summarise a free-text customer requirement, classify it against a known set of implementation patterns, draft a suggested next step and flag uncertainty for a human. The surrounding workflow can still be automated. The approvals can still be explicit. The data passed into the model can still be limited to what the task actually needs.</p><p>This is the difference between using AI as a magic layer and using AI as a component. The magic layer approach is tempting because it looks impressive in a demo. The component approach is harder because it requires architecture, measurement and restraint. But the component approach is the one that survives contact with real operations.</p><p>The same applies at an organisational level. AI adoption should not be measured only by how many employees have access to a tool or how many prompts they send in a month. Those numbers may be useful, but they are not value. Value is time removed from a process, error rates reduced, customer wait times shortened, decisions improved, risks surfaced earlier, knowledge made easier to reuse or revenue unlocked through something that was previously impossible.</p><p>That is why I think the next phase of AI adoption will be less glamorous and far more interesting. The first phase was experimentation. The second phase was rollout. The next phase will be operationalisation. It will involve finance teams asking harder questions about variable spend. Security teams asking harder questions about data exposure. Operations teams asking harder questions about process ownership. Technology teams asking harder questions about architecture, observability and failure modes. Business leaders asking whether a use case is genuinely worth scaling or whether it was only ever a good demo.</p><p>None of this means we should be cynical about AI. I use AI constantly, and I think it is one of the most important technologies of our lifetime. It can accelerate research, improve writing, help engineers reason through unfamiliar code, support customer teams, summarise complex information, generate ideas, expose patterns and make many kinds of knowledge work faster. Used well, it can be extraordinary.</p><p>But “used well” is doing a lot of work in that sentence.</p><p>Using AI well does not mean using it indiscriminately. It does not mean replacing every script with an agent or every workflow with a prompt. It does not mean pretending deterministic problems are suddenly reasoning problems because the market has decided that AI is the new badge of innovation. It means understanding the shape of the work and choosing the right mechanism for each part of it.</p><p>Sometimes that mechanism will be a large language model. Sometimes it will be a smaller model. Sometimes it will be retrieval. Sometimes it will be a rules engine. Sometimes it will be a scheduled job, a script, a queue, a form validation, a dashboard, a workflow tool or a simple reminder. The maturity is not in choosing the most advanced option. The maturity is in choosing the most appropriate one.</p><p>That may become a competitive advantage sooner than people think. As AI pricing becomes more closely tied to consumption, organisations that understand their processes will have better cost control. Organisations that manage context carefully will have more predictable systems. Organisations that separate automation from reasoning will build workflows that are easier to govern. Organisations that know when not to use AI will waste less money than those trying to prove they are using it everywhere.</p><p>The uncomfortable truth is that many businesses do not have an AI strategy problem. They have a process understanding problem that AI has made more visible. AI can help solve that, but it can also disguise it for a while. The demo will work. The pilot will impress people. The prototype will generate excitement. Then the real questions will arrive. Who owns this process? What happens when the model is wrong? What does it cost at scale? What data does it need? What should be automated before AI is involved? What should remain human? What does success actually mean?</p><p>Those are not anti-AI questions. They are operational questions. They are the questions that turn AI from theatre into infrastructure.</p><p>So, should you use AI in everything you do?</p><p>Not quite.</p><p>You should consider AI in everything you do. You should understand where it could help, where it could change the economics of a process and where it could create new possibilities. But you should not assume that every problem deserves an AI-shaped answer.</p><p>The goal was never to use AI everywhere. The goal was always to solve problems well.</p><p>And sometimes the smartest solution is still the simplest one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1698e632a0b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Skills Aren’t New. Here’s What’s Actually Happening Inside an LLM.]]></title>
            <link>https://olivertappin.medium.com/claude-skills-arent-new-here-s-what-s-actually-happening-inside-an-llm-52df387ba18e?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/52df387ba18e</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[claude-skills]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Wed, 06 May 2026 19:26:18 GMT</pubDate>
            <atom:updated>2026-05-06T20:05:42.373Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ovij2pcwOd_uutx1" /><figcaption>Photo by <a href="https://unsplash.com/@jjying?utm_source=medium&amp;utm_medium=referral">JJ Ying</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>The hype cycle is doing what hype cycles always do: inventing names for things that already have names.</h4><p>That isn’t meant to be dismissive. The patterns are real and they work. The point is that you don’t need new mental models to understand most of what’s going on in this space. You need to recognise that the same primitives you’ve been using to build software for years — indexes, working memory, isolated execution — are what these systems are running on. If you’re a software engineer feeling lost in the AI conversation, you already have the vocabulary. You’re just not being told it’s the same vocabulary.</p><p>But — and this is the bit that genuinely matters — there is a layer underneath where the analogies break. Where Skills are not indexes, context is not RAM, and a subagent is not a CI container. That layer is where the actual model lives, and understanding it is what separates people who can prompt-engineer from people who can build production systems.</p><p>So I want to do two things in this post. First, defend the analogies and tell you to use them, because they get you 80% of the way there for free. Then, go underneath and show you the 20% the analogies miss, because that’s where the real engineering problems live.</p><h4>Skills are indexes (mostly)</h4><p>Anthropic Skills are the latest naming convention for “small markdown file describes a capability, model reads it on demand, model executes.” If that sounds like a primary key pointing to a row in a table, you’re not wrong. The mental model holds. You have an index — the SKILL.md frontmatter, name plus description — that the model scans first to figure out what’s available. When something matches the request, the model fetches the body, which is the row data. The body might reference further files, which is your foreign-key join. You can build the same thing yourself with a top-level INDEX.md and discipline. The pattern is “small index, large lazy-loaded modules” and operating systems have done this since the 1960s.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CBmPpYJ2kX-5Ph99KECm7w.jpeg" /><figcaption>A SKILL.md file must begin with YAML Frontmatter that contains a file name and description, which is loaded into its system prompt at startup.</figcaption></figure><p>So treat Skills as an index. That’s a useful, accurate frame. Most of what’s written about them is overcomplicated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/1*QbLncM0OxXwM0HKTI3v1Cg.png" /><figcaption><a href="https://github.com/anthropics/skills/tree/main/skills">https://github.com/anthropics/skills/tree/main/skills</a></figcaption></figure><p>Where it stops being a database index is in the lookup mechanism. A B-tree gives you O(log n) deterministic retrieval. A hash table gives you O(1) deterministic retrieval. The model’s “lookup” of which skill to call is neither. It’s a forward pass through billions of parameters where attention over the natural-language descriptions determines which skill the model believes is relevant to the current request. The match is probabilistic. It’s closer to vector retrieval over semantic embeddings than to a B-tree traversal. Two skills with overlapping descriptions can confuse the model. A skill with a vague description can be silently ignored. The “index” doesn’t have a deterministic correctness guarantee; it has a confidence distribution.</p><p>This matters in production. If you treat Skills as a deterministic index, you’ll be surprised when the model picks the wrong one or misses the right one. If you treat them as a probabilistic retrieval layer over natural language — which is what they actually are — you start writing descriptions the way you’d write embeddings: distinct, specific, non-overlapping, designed to maximise the discriminative signal in the model’s attention pattern. That’s a different engineering discipline. The analogy gets you to the door. The mechanism gets you inside.</p><h4>Context windows are RAM (mostly)</h4><p>This one I see all the time. Context window equals working memory: expensive, limited, fast access. Externalise to disk when you don’t need it. As a coarse capacity model, that’s fine. Use it for back-of-envelope reasoning. You have N tokens, each turn costs you compute proportional to those tokens, treat them as a budget.</p><p>The analogy breaks the moment you ask about access patterns. RAM is uniformly addressable. Every byte takes the same time to fetch. Context is not. There’s a well-documented effect called “lost in the middle” — tokens at the start and end of context are attended to more reliably than tokens buried in the middle, even when those middle tokens are perfectly relevant. This is partly architectural — modern transformers use rotary positional encodings (RoPE) which have their own attenuation properties at long range — and partly distributional, because training data has structural patterns (instructions at the top, conclusions at the end) that the model internalises. Position is not neutral. The middle of your context is genuinely worse real estate than the edges.</p><p>The deeper break is dilution. RAM doesn’t degrade when you fill it; what’s already loaded retrieves at the same speed regardless of what else is loaded. Context does degrade, long before you hit the limit. Here’s why: attention is a softmax over all previous tokens. Every token’s attention weights sum to one. So if you have one critically relevant token in a 200k-token context window, it’s competing with 199,999 other tokens for that probability mass. Add more irrelevant tokens and the relevant one’s relative weight goes down. The signal-to-noise ratio in the attention scores degrades. The model becomes more likely to surface plausible-but-wrong content, because the relevant content has less relative weight in the softmax distribution.</p><p>This is why “stuff more context in” isn’t a free improvement. It actively harms what’s already there. The closest hardware analogy isn’t RAM — it’s a CPU cache where every fetch is a fuzzy similarity search and the cache lines compete with each other for relevance. RAM doesn’t work like that. Caches do, partially. Attention is more aggressive than either.</p><p>The practical implication: you don’t optimise context the way you optimise memory. You don’t ask “how much can I fit.” You ask “what’s the smallest, cleanest set of tokens I can leave in context to do this turn well, and where do I position them?” That’s a different optimisation problem. RAM gets you to the door. The mechanism gets you the answer.</p><h4>Subagents are CI containers (mostly)</h4><p>Spawn an isolated environment with a defined task, fresh state, controlled inputs, get a result back. Subagent invocations work this way and the analogy is genuinely useful. Process isolation is process isolation. If you’ve built a CI pipeline, you already understand the orchestration model: parent dispatches a subtask to a child with a clean working space, child completes the work, child returns its output, parent doesn’t carry the working memory of the child. The whole point is that the parent context stays lean.<br>I actually like this one as a teaching analogy because it generalises well.</p><p>You can reason about subagent failure modes the same way you reason about CI failure modes: timeouts, resource limits, undefined behaviour at boundaries, output validation. It’s not a perfect map but it’s productive.</p><p>Where it breaks is determinism. A CI container, given the same inputs, produces the same outputs. (Caveat: flaky tests exist. But that’s considered a bug.) A subagent invocation is a stochastic forward pass. Same inputs, same prompt, same model, even at temperature zero, you can still get different outputs across runs because of non-deterministic floating-point reduction order on the GPU. The “container” passes integration tests probabilistically, not deterministically. The contract is statistical.</p><p>The implication for production: you can’t treat subagent outputs the way you treat function returns. You need verification, retries, structured output validation, sometimes consensus across multiple invocations. The orchestration patterns from CI mostly transfer — but the assumption that “the unit of work either succeeds or fails cleanly” doesn’t. Sometimes the subagent succeeds-with-a-subtly-wrong-answer, which CI containers don’t really do. That’s a class of failure mode requiring its own engineering response, and the analogy doesn’t prepare you for it.</p><h4>What’s actually new</h4><p>If most of the patterns aren’t new, what is? In my view, exactly one thing: attention as a soft retrieval mechanism over a finite, competitive resource. That’s the primitive. Everything else — Skills, RAG, subagents, context engineering — is a way of shaping the inputs to that mechanism so it produces useful outputs more often.</p><p>You can hold the whole framework together with three observations that fall out of the architecture:</p><p>The model has no memory between turns. Every turn is a fresh forward pass over the entire context. Anything you want preserved must be in tokens this turn — directly, or fetchable via tool call. State lives in your harness, not in the model. This is a brutal constraint compared to the systems we usually build, where state persists by default.</p><p>Attention is a fixed budget split across everything in context. Adding tokens is not free. Irrelevant content harms relevant content’s retrieval probability. The right question is not “how much can I fit” but “what’s the minimum I need.”</p><p>Position matters. The middle of context is the worst place to put critical content. Treat the start and end as premium real estate. Treat the middle as the place stuff goes to be partially forgotten.</p><p>Once you internalise these three, every “context engineering technique” becomes legible as a way to increase signal-to-noise in the working set. Skills do this by keeping bodies out of context until needed. RAG does this by retrieving only top-k matches. Subagent delegation does this by giving the parent only the final output, not the working scratchpad. Anchor positioning does this by putting critical content at the edges. None of these are new techniques — externalisation, retrieval, process isolation, and good prompt structuring are all decades old. What’s new is that they now compose around an attention mechanism, not a database.</p><h4>Stick to what you know</h4><p>So here’s where I land: stick to the software engineering primitives you already understand. They explain most of what you need. Skills are indexes. Context is RAM-ish. Subagents are containers. Use those frames. They won’t lead you wrong at the level most teams are working at.</p><p>But know where the analogies stop. The lookup is probabilistic, not deterministic. Working memory is competitive, not uniform. Isolation is statistical, not pure. Those three breakdowns are where the engineering gets interesting, and where the people building production AI systems actually spend their time. If you’re a leader making decisions in this space, that’s the layer you want your team operating at — not the marketing layer, not the analogy layer, but the layer underneath where the model actually lives.</p><p>The hype is busy reinventing names. The work is the same as it’s always been. Understand your primitives. Know where they break. Build accordingly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52df387ba18e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Handover Hell: The Controversial Truth Behind Workplace Inefficiency]]></title>
            <link>https://olivertappin.medium.com/handovers-dont-work-here-s-why-d55c4203752a?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/d55c4203752a</guid>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[documentation]]></category>
            <category><![CDATA[scrum]]></category>
            <category><![CDATA[handover]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Sat, 23 Dec 2023 11:42:32 GMT</pubDate>
            <atom:updated>2024-07-30T11:05:32.158Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_dN7mDA7R4ZQHb47" /><figcaption>Photo by <a href="https://unsplash.com/@mimithian?utm_source=medium&amp;utm_medium=referral">Mimi Thian</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Time and time again, businesses all over the world are spending valuable hours sharing their knowledge about a particular project with others, to show them how it works, all the caveats that come with it. Sometimes, the handover shows usage which may not even be the most efficient, or it might be done for a particular reason, and that person may not have asked at the time ‘why do we do it in this way?’, but they’ll carry on doing it. Maybe even after that person has left the company.</p><p>The problem is, if you have to explain that to someone, in that much detail, about how something works, then you’ve done something <em>seriously</em> wrong.</p><p>This becomes the result of not spending enough time documenting, in clear language, what is required from a new person, to be able to perform a particular task, or show someone how a particular project or library works. If you’re not putting yourself the the shoes of someone who’s never seen this project before, then you’re setting them up to fail.</p><p>In an ideal world, you want someone that can come along and just crack on with what’s required, rather than having to spend an hour or two, of someone’s valuable time, telling them all the ins-and-outs of that particular project and how to use it.</p><p>If we take a moment to think about how the open-source world works, where there are <em>millions</em> of people online, and you want anyone to come along and to be able to use your project or vendor in a way where you don’t have to have a phone call with them for an hour to explain to them how it works you just want them to get on with it and have all the information that they need there. If you take that same mentality of the open-source world onto a business, and everyone has documentation to read from, then things become a thousand times easier, because you don’t need to spend time money on people telling other people how to do something.</p><p>The cycle of repetitive handovers is not only inefficient but can also be a significant drain on resources. Imagine spending upwards of 30 hours explaining the same project to multiple individuals — it’s a scenario that highlights the inefficiency and time-consuming nature of poor knowledge transfer. The solution lies in creating comprehensive, accessible documentation. With such a resource, it doesn’t matter if you’re unavailable due to vacation, illness, or other commitments. Anyone can refer to the document, understand the project requirements, and take the necessary actions. This approach ensures continuity, saves time, and guarantees that tasks are completed effectively, regardless of who is available.</p><p>To clarify, this doesn’t mean you should stop communicating with your team about project requirements or forego social interactions when they are beneficial. These aspects are vital for team dynamics and project success. However, equipping your team with comprehensive tools and resources allows them to progress in their work autonomously, without being dependent on others for information. This approach prevents bottlenecks caused by unavailability of key personnel, ensuring smooth workflow and consistent productivity.</p><p>However, this raises an important question: How can organisations effectively create and maintain such documentation? The answer lies in a systematic approach to knowledge management.</p><p>Firstly, organisations must recognise the importance of documentation as a core part of their operational strategy. This means allocating time and resources to document processes, projects, and systems. It’s not just about writing down steps; it’s about creating a comprehensive, easy-to-understand guide that can be used by anyone, regardless of their familiarity with the project.</p><p>This leads to the second point: the documentation must be accessible and user-friendly. Complex jargon, technical terms, and convoluted explanations can be more of a hindrance than a help. The goal is to simplify, not complicate. This might involve using visuals, step-by-step guides, FAQs, and even video tutorials. Remember, the aim is to make the onboarding process as smooth as possible.</p><p>Thirdly, documentation should be a living document. It needs regular updates and revisions to stay relevant. As projects evolve, so too should the documentation. This requires a culture where updating documentation is seen as an integral part of project development, not an afterthought or a burdensome task.</p><p>But why is this so crucial? The benefits are manifold. For one, it significantly reduces the time spent on handovers. New team members can hit the ground running, delving into the documentation to understand the project’s ins and outs. This not only speeds up the onboarding process but also enhances the overall productivity of the team.</p><p>Moreover, good documentation can serve as a safety net. In instances where key team members are unavailable, others can step in without the project coming to a standstill. This continuity is vital in today’s fast-paced business environment where delays can have significant repercussions.</p><p>Furthermore, having a well-documented process aids in maintaining consistency and quality. When everyone follows the same guidelines and procedures, there’s less room for error and variation in output. This consistency is particularly crucial when dealing with complex projects or when working in regulated industries.</p><p>It’s also worth noting that comprehensive documentation can be a valuable tool for training and development. New employees can learn a lot about the company’s processes and projects through these documents, which can accelerate their learning curve and integration into the team.</p><p>While handovers might seem like a necessary part of business operations, they often signify a deeper issue with knowledge transfer within an organisation. By investing in thorough, accessible, and regularly updated documentation, businesses can improve efficiency, maintain continuity, ensure consistency, and foster a more autonomous and empowered workforce.</p><p>The goal is not to eliminate communication and collaboration but to make it more effective and efficient, ensuring that team interactions are value-adding and not just about filling in knowledge gaps.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d55c4203752a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sustainable Coding: Building Eco-friendly and Efficient Software]]></title>
            <link>https://olivertappin.medium.com/sustainable-coding-building-eco-friendly-and-efficient-software-84227c9da4f5?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/84227c9da4f5</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[sustainable-development]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Fri, 22 Dec 2023 18:35:22 GMT</pubDate>
            <atom:updated>2023-12-22T18:35:22.331Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_2fdTouZXZqQ_V1Z" /><figcaption>Photo by <a href="https://unsplash.com/@nasa?utm_source=medium&amp;utm_medium=referral">NASA</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In an era where environmental sustainability is a global priority, the field of software development is not immune to the call for greener practices. As software engineers, we have the unique opportunity to influence the environmental footprint of our digital creations. This article explores how to make software development more eco-friendly, focusing on efficient coding practices, sustainable infrastructure, and understanding the environmental impact of technology.</p><h4>1. Efficient Coding Practices for Sustainability</h4><p>Efficient coding is not just about optimising performance; it’s also about reducing the environmental impact. Minimally resource-intensive software requires less energy to run, translating to a smaller carbon footprint.</p><ul><li>Code Optimisation: Efficient algorithms use fewer computational resources. Optimise code to reduce unnecessary CPU cycles and memory usage, which in turn lowers energy consumption.</li><li>Resource Management: Efficiently managing application resources like memory and connections can significantly reduce a software’s environmental impact.</li><li>Green Coding Standards: Adopt coding standards focused on sustainability. These standards can include guidelines for energy-efficient coding, resource utilisation, and waste reduction.</li></ul><h4><strong>2. Sustainable Infrastructure and Cloud Services</strong></h4><p>The choice of infrastructure plays a crucial role in the sustainability of software. Cloud services have been at the forefront of this, with major providers investing heavily in green initiatives.</p><ul><li>Green Cloud Providers: Google Cloud has been a leader in sustainability, <a href="https://cloud.google.com/sustainability">claiming</a> to match 100% of their energy consumption with renewable energy since 2017. Their commitment to running on carbon-free energy 24/7 by 2030 sets a high standard for other providers.</li><li>Energy-Efficient Data Centres: Opt for cloud services that use energy-efficient data centres. This includes efficient cooling systems, sustainable energy sources, and carbon-neutral operations.</li><li>Serverless Architecture: Serverless computing can be more energy-efficient as it allows for precise scaling, ensuring that only the necessary resources are used.</li></ul><h4><strong>3. Minimising the Environmental Impact of Technology</strong></h4><p>Understanding the broader impact of our technological choices is key to sustainable software development.</p><ul><li>Lifecycle Assessment: Consider the environmental impact of software throughout its lifecycle, from development and deployment to maintenance and decommissioning.</li><li>Data Transfer and Storage Optimisation: Reducing data transfer and optimising storage can lower the energy consumption of data centres.</li><li>E-Waste Management: Promote practices that reduce e-waste, like extending the lifespan of hardware and responsibly recycling electronic components.</li></ul><h4><strong>4. The Role of Developers in Promoting Sustainability</strong></h4><p>As developers, we play a vital role in advocating for and implementing sustainable practices.</p><ul><li>Education and Awareness: Stay informed about the latest in sustainable technology and share knowledge with peers.</li><li>Advocacy for Green Policies: Advocate for environmental considerations in your organisation’s IT policies.</li><li>Community Involvement: Participate in or contribute to open-source projects that focus on sustainability in technology.</li></ul><p>Sustainable coding and the choice of eco-friendly infrastructure are crucial in reducing the environmental impact of our digital world. As software engineers, we are in a unique position to lead the charge in this vital effort. By adopting efficient coding practices, opting for sustainable infrastructure, and understanding the broader environmental impact of our technology choices, we can make a significant difference in building a more sustainable future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=84227c9da4f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Embracing the Future: How AI is Transforming Software Development]]></title>
            <link>https://olivertappin.medium.com/embracing-the-future-how-ai-is-transforming-software-development-db5e6f3223ad?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/db5e6f3223ad</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[github-copilot]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Thu, 21 Dec 2023 17:21:57 GMT</pubDate>
            <atom:updated>2023-12-21T17:21:57.738Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JrWVo67TJiNWRYKL" /><figcaption>Photo by <a href="https://unsplash.com/@thisisengineering?utm_source=medium&amp;utm_medium=referral">ThisisEngineering RAEng</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The emergence of Artificial Intelligence (AI) has revolutionised numerous industries, with software development being no exception. This article delves into how AI is reshaping the field of software engineering, enhancing efficiency, and opening new avenues for innovation.</p><h4><em>AI in Code Generation and Assistance</em></h4><p>Tools like GitHub Copilot represent a significant leap in coding efficiency. By suggesting code snippets and entire functions, these AI assistants allow developers to focus on more complex and creative aspects of their projects. Moreover, AI’s capability to analyse vast codebases leads to more robust and error-free code.</p><h4><em>AI in Debugging and Predictive Analysis</em></h4><p>One of the more groundbreaking applications of AI in software development is in predictive analysis and debugging. AI algorithms can anticipate potential issues in the coding phase, saving significant time and resources in the development cycle. Automated testing, powered by AI, has become a game-changer, enabling developers to identify and resolve bugs more efficiently than ever before.</p><h4><em>Balancing AI and Human Creativity</em></h4><p>While AI’s contributions are immense, it’s crucial to remember the irreplaceable value of human creativity and intuition in software development. AI tools should be seen as complements to human skills, assisting rather than replacing the developer’s unique problem-solving abilities.</p><h4><em>Ethical Considerations and Best Practices</em></h4><p>The integration of AI in software development raises several ethical considerations, such as data privacy and bias in AI algorithms. It’s essential for developers to adhere to best practices that prioritise ethical considerations and ensure responsible use of AI technologies. My personal recommendation would be to read the terms and conditions, understand the policies and do not enter in any personally identifiable information (PII) into any system.</p><p>AI is undeniably transforming software development, making processes more efficient and opening new horizons for innovation. Although it’s not quite there yet, especially to start replacing anyone, it will of course make us more efficient. Continue to embrace these changes, stay informed and adapt. This is key to leveraging AI’s full potential in software engineering.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db5e6f3223ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[4 Reasons Why You Need Automation]]></title>
            <link>https://olivertappin.medium.com/4-reasons-why-you-need-automation-e54d2f8061ca?source=rss-5eca2b78a00b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e54d2f8061ca</guid>
            <category><![CDATA[time-management]]></category>
            <category><![CDATA[scalability]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Oliver Tappin]]></dc:creator>
            <pubDate>Wed, 25 Nov 2020 13:20:48 GMT</pubDate>
            <atom:updated>2020-11-25T13:20:48.202Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xnUTxFkGBk1CwG-x" /><figcaption>Photo by <a href="https://unsplash.com/@_imkiran?utm_source=medium&amp;utm_medium=referral">Sai Kiran Anagani</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Like most Software Engineers, I started my career by being interested in solving problems. It took me a while to realise that some of those tasks were daily issues. By nature, we want to solve that problem fast and move on to the next thing, and this might mean using the quickest and dirtiest method available. But I knew, deep down, solving problems using that method wasn’t scaleable. It was just firefighting.</p><p>I have a vision to help create a world in which we can do <em>more</em> using the time that we have available. Part of being able to achieve that is writing code that will solve problems without human interaction.</p><h3>Remove unnecessary human error</h3><p>If you’ve ever spoken to a Site Reliability Engineer, they’ll tell you that most causes of outages are due to human error. After all, a machine is only designed to do what a human has programmed it to do — and we can’t get away from the fact that we are all fallibly human.</p><p>We’ve all been there. Ask anyone who’s worked in the industry long enough and they’ll happily share some of their most painful experiences. Whether it’s dropping a table in a production database or removing all healthy instances from a load balancer, each story will stay with them forever.</p><p>Most of the time, those mistakes don’t lead to disciplinary action. They’re taken as a positive learning experience and provide a platform for that department to improve. That’s why we have Root Cause Analysis (RCA) documents or Autopsy Reports, so that we can learn how the situation occurred. That’s also why there are plenty of tools out there that mimic those kind of behaviours, such as <a href="https://github.com/Netflix/chaosmonkey">Chaos Monkey</a>, created by Netflix.</p><h3>Because time waits for no man (or woman)</h3><p>Just imagine if you had the ability to carry out the mundane tasks in your daily routine, without you having to be there, and you could spend that time doing something else — something exciting or something really <em>worthwhile — </em>or maybe just nothing at all. The choice would be yours.</p><p>This led me to the idea of automating everything that I possibly could. If I was doing a task that took longer than 5 minutes that I knew I would have to repeat, I would try to find a way to automate it. I was fascinated with being able to code, seeing that code translate into tasks and having the satisfaction of knowing I had created a useful tool to save time and effort later.</p><p>The idea of creating all these efficiencies within your day can free up that time for you to do something else is very attractive. Doing the same mundane tasks day in and day out doesn’t get you anywhere. It doesn’t use your brain, your creativity or your aptitudes.</p><h3>Using time wisely frees it up to do what you love</h3><p>If you could save 30 minutes a day, from 10 members of your team, based on an 8 hour day and 260 working days a year, that’s 162.5 days you’ve got back. Imagine having that extra time on your hands. What <em>else</em> could you do?</p><p>Several years ago, I started a new job at a new Digital Agency in London. They were growing rapidly and had some large and interesting projects on the go. I took the job because it seemed exciting, and fulfilled the overwhelming drive I had to ‘get stuff done’. If you’ve worked at an agency before, it’s a bit like getting on a moving train. Everything needs to be done yesterday and you don’t have time to do what you had in your head. Instead, you end up with some bastardised version that just about works and at the end of the day little idea of what you’ve achieved.</p><p>This agency mindset allowed me to think of an abstract way to automate as much as I could, so that when a new project came along, I could focus on the elements that I <em>wanted to do; the things which caught and fired up my imagination. </em>If you can identify the repetitive tasks, then you can set yourself up for automation.</p><h3>Learn from your corrections</h3><p>The old saying goes: “Learn from your mistakes”, but making mistakes doesn’t really do much other than identify where things have gone wrong. Correcting those mistakes is where the value lies, and the key to people expanding their knowledge and progressing.</p><p>It’s very easy for someone knowledgeable to step in and do something, hiding the reality of manual changes, without documentation or sharing it with other members of the team. I always try to put myself in the shoes of someone who has never deployed a particular application, or has never seen this side of the code before, and try to share as much as possible. Knowledge is power and it should be shared: there is no room for selfishness and the more altruistic you are about this, the more you will find team members cooperating, helping each other and dropping the blame mentality when things go wrong: they will all be eager to help work out how to fix the problem. Automating the mundane can help build a cooperative and solid team.</p><p>Try and automate as much as possible, where it makes sense. Never catch yourself doing the same repetitive task ever again. Life’s too short.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e54d2f8061ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>