<?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 Anvayasharma on Medium]]></title>
        <description><![CDATA[Stories by Anvayasharma on Medium]]></description>
        <link>https://medium.com/@anvayasharma509?source=rss-4dc956b62639------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jaVOh1l066BenLx-GqlWzg.jpeg</url>
            <title>Stories by Anvayasharma on Medium</title>
            <link>https://medium.com/@anvayasharma509?source=rss-4dc956b62639------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 05:24:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anvayasharma509/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[The Best Expense Tracker Is the One You’ll Actually Use]]></title>
            <link>https://medium.com/@anvayasharma509/the-best-expense-tracker-is-the-one-youll-actually-use-b43b6fdbe2ac?source=rss-4dc956b62639------2</link>
            <guid isPermaLink="false">https://medium.com/p/b43b6fdbe2ac</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[personal-finance]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Anvayasharma]]></dc:creator>
            <pubDate>Fri, 11 Sep 2026 09:13:59 GMT</pubDate>
            <atom:updated>2026-09-11T09:13:59.752Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aBAQEo2vvP1iZln4_VbcBw.png" /></figure><p>I have tried every reasonable way to track my expenses. Dedicated budgeting apps. Notion templates. Spreadsheets with color-coded categories and monthly formulas. Each attempt followed the same arc: two weeks of diligent logging, a few days of guilt-skipping, and then quiet abandonment.</p><p>The problem was never the tool. The tools were fine. The problem was that every one of them made the moment of logging feel like a chore. Open the app. Tap the plus button. Pick a category from a dropdown. Type the amount. Add a note. Save. That’s six steps for something that should feel like a reflex.</p><p>It took me an embarrassingly long time to realize that this wasn’t a discipline problem. It was a design problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BwDfBU8aneGloRYQFWOmKg.png" /></figure><h3>The only moment that matters</h3><p>Most expense tracking products compete on what happens after you’ve logged your data. Better charts. Smarter budgets. Prettier breakdowns by category. But none of that matters if the data never gets entered in the first place.</p><p>The bottleneck isn’t analysis. It’s capture.</p><p>And the result of high-friction capture is predictable. You log the big stuff (rent, flights, subscriptions) and skip the small stuff (coffee, auto, snacks). Your data is structurally incomplete, and the entire system slowly becomes useless. Not because it broke, but because you stopped feeding it.</p><p>I wanted logging an expense to feel closer to sending a text than filling out a form. And I wanted my data to land somewhere I already knew how to read: a spreadsheet.</p><h3>What I built</h3><p>I built a Telegram bot. You message it like you’d message a person, and it turns your message into a structured row in a Google Sheet.</p><p>You type: flight 8977 delhi trip</p><p>The bot replies: Logged: Flight | 8977 | Delhi trip | 25 Jun</p><p>No app to open. No form to fill. No category to select. If you mention a date (“june 28”), it logs to that date. If you don’t, it defaults to today. Each month gets its own tab in the sheet, created automatically.</p><p>Under the hood, the bot sends every message to a language model for parsing. The model reads your freeform text and extracts four things: category, amount, note, and date. It returns a small structured object, and the bot writes that as a row to your sheet. This means there’s no fixed format to memorize. “lunch 450,” “450 lunch,” “had lunch for 450 near office” all resolve to the same entry. You don’t adapt to the tool. The tool adapts to you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vOAGUNXupHdNYRqwvjSvgw.png" /></figure><p>The data lands in Google Sheets because that’s where I was already comfortable looking at my finances. You can open your sheet anytime, see every entry, edit anything, add your own formulas. The sheet is the entire review UI, and it’s a UI you already know.</p><p>Queries work the same way. Type “total this month” and the bot tells you. “Show flights” lists every flight entry. For open-ended questions, the bot reads your recent data and answers in plain text. No dashboards. Just ask.</p><p>The whole thing runs on a single webhook endpoint. Telegram pushes your message to a small server, the server calls the language model, writes to your sheet, and replies. No session state, no database, no background processes. The architecture is almost comically simple, and that’s the point.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YIr3oKu9SZha7bgAB06cmw.png" /></figure><h3>The decisions that made it stick</h3><p>What I want to talk about isn’t the bot itself. It’s the tradeoffs behind it. Each one goes against conventional product logic, and each one is why this tool stuck where others didn’t.</p><p><strong>I put it in a chat, not an app.</strong> The most underrated quality of a chat interface is that it’s already open. I’m already in Telegram talking to people. The expense bot is just another conversation in my list. The context switch cost is zero. A dedicated app means leaving what you’re doing, finding the app, waiting for it to load, and starting a workflow. By the time you’ve done all that, the moment has passed.</p><p><strong>I used a language model instead of a form.</strong> A traditional app gives you fields: amount, category, date, notes. A form is unambiguous, but it’s also slow and rigid. The language model lets me type the way I think. I don’t have to decompose my expense into four fields. I just describe it the way I’d tell a friend: “auto 250 to office.” The cognitive cost of translating a thought into a form is real, and it’s the kind of micro-friction that kills daily habits.</p><p><strong>I stored data in a spreadsheet, not a database.</strong> This is the decision people question most. But a database requires a UI on top of it, and I didn’t want to build one. Google Sheets is the UI. I was already comfortable tracking finances in a spreadsheet. The bot just removed the part I hated: manually typing rows. The data format didn’t change. The entry method did.</p><p><strong>I left out budgeting, charts, and alerts.</strong> Every expense app ships with budgeting features. I deliberately didn’t. Not because budgeting is useless, but because it’s the wrong thing to optimize first. If the tool can’t get me to log consistently, no amount of budgeting logic matters. I wanted the product to do one thing well: capture expenses at the speed of thought.</p><h3>What changed when friction disappeared</h3><p>When logging takes three seconds and requires no app switch, you log everything. The 30-rupee chai. The 15-rupee parking. The random impulse buy at a street stall. Things I would never have opened an app to record.</p><p>The result is that my data is complete for the first time. Not approximately complete. Actually complete. And complete data changes what expense tracking can do for you. When every transaction is there, your monthly total is real. Your category breakdown is honest. You can actually see where the money goes, not just where the big money goes.</p><p>The habit changed too. I don’t think about logging anymore. It’s not a task I remember to do. It’s just part of the act of spending. Buy something, text the bot. The friction was the only thing standing between me and consistency, and once it was gone, the behavior locked in on its own.</p><p>The bot has no charts, no budget alerts, no multi-currency support. I’m not in a hurry to close those gaps. Every feature I add could make the core interaction heavier. The bot works because it does one thing instantly. Features will be added with one filter: does this make capture faster or slower? If slower, it doesn’t ship.</p><h3>The broader point</h3><p>Most productivity software asks you to change your habits to match the product. Open this app. Follow this workflow. Adopt this system. That didn’t work for me.</p><p>What worked was meeting myself where I already was: in a chat window, comfortable with spreadsheets, unwilling to fill out forms. The bot didn’t ask me to become more disciplined. It just removed the thing that was making discipline necessary.</p><p>If you’re building a personal tool and people aren’t using it, the answer probably isn’t more features. It’s less friction. Find the moment of capture, and make it disappear.</p><p>PS:</p><p>The bot is open source. If you want to set up your own, the repo is [<a href="https://github.com/anvayasharma509-svg/whatsapp-expense-logger">here</a>]</p><p>How to Setup?<br>Link: Guide in Project</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b43b6fdbe2ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Are So Delusional]]></title>
            <link>https://medium.com/@anvayasharma509/we-are-so-delusional-688ece854674?source=rss-4dc956b62639------2</link>
            <guid isPermaLink="false">https://medium.com/p/688ece854674</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[books]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Anvayasharma]]></dc:creator>
            <pubDate>Sat, 15 Aug 2026 21:40:12 GMT</pubDate>
            <atom:updated>2026-08-15T21:40:12.469Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong><em>On Factfulness, the gap between perception and reality, and the decisions that shape all of us</em></strong></h4><p>I was reading a by Hans Rosling book when it hit me.</p><p>He had spent years giving a simple quiz to some of the most powerful people in the world. Davos attendees. Nobel laureates. UN officials. Twelve questions about basic global facts. Things their decisions depended on.</p><p>They scored worse than chimpanzees choosing randomly.</p><p>Not slightly worse. Consistently, reliably worse. A chimp picking answers by chance would outscore the people shaping global policy on the very issues they’d built their careers around.</p><p>My reaction was not academic: <em>Oh damn. We are so delusional.</em></p><p>The book is Factfulness. And that finding is not even the most important thing in it.</p><p>The most important thing is what Rosling found when he asked why.</p><p>The errors weren’t random. If people were simply uninformed, their wrong answers would scatter in all directions. They didn’t. Every error pointed the same way: toward a world that was more frightening, more divided, more hopeless than the data showed. People weren’t missing information. They were replacing it with something else.</p><p>Rosling called them dramatic instincts. He identified ten.</p><p>The <strong><em>gap instinct </em></strong>divides every situation into two extremes, erasing the messy middle where most of reality lives. The <strong><em>negativity instinct </em></strong>makes bad news feel more true than good news, regardless of evidence. The <strong><em>straight line </em></strong>instinct assumes any trend moving in one direction will keep moving that way forever. The <strong><em>urgency instinct </em></strong>is perhaps the most dangerous: it shuts down careful thinking the moment it fires, replacing analysis with the overwhelming feeling that something must be done right now.</p><p>These aren’t signs of stupidity. They’re features of human cognition. Useful in some situations. Deeply misleading in others.</p><p>The problem Rosling identified is not that humans have these instincts. It’s that humans with these instincts are the ones making the decisions.</p><p>I’m early in my career. I sit near rooms where decisions get made. And reading this book, I kept having the same quiet thought: this is not about global development statistics. This is about every room where something consequential is being decided.</p><p>The gap Rosling documented doesn’t close when you change the domain. It follows the decision-maker. Into boardrooms. Into strategy meetings. Into technology bets and hiring cycles and market assumptions that ripple outward in ways invisible at the source but felt everywhere downstream.</p><p>What unsettled me most wasn’t that powerful people are wrong. It was that they’re wrong in a predictable way, driven by machinery they can’t see, producing consequences they don’t trace.</p><p>Rosling wasn’t angry at the Davos attendees who failed his quiz. He was troubled by something simpler: their decisions, made on distorted mental models, shaped the lives of people who had no seat at the table.</p><p>The delusion isn’t confined to the person holding it. It travels with every decision they make.</p><p>Nowhere is this more visible right now than in business and technology.</p><p>This isn’t an argument that big companies are reckless, or that the people running them are foolish. Google, Meta, Amazon, Microsoft, Apple: these are among the most analytically sophisticated organizations ever built. More data, more expertise, more institutional knowledge than almost any decision-making body in history.</p><p>And they still make calls on distorted perception. Because distorted perception is not a resource problem. It’s a human one.</p><p>When companies at this scale make decisions, the consequences don’t stay inside the building. A strategic bet made on a trend that doesn’t hold doesn’t just cost the company money. It shapes what thousands of smaller companies do next. It moves labor markets. It determines what gets built, what gets funded, who gets hired, what problems get attention and which ones stay invisible.</p><p>The people in those rooms aren’t making decisions carelessly. They’re making them humanly. Which, Rosling showed us, is a different thing from making them accurately.</p><p>The urgency instinct fires and analysis compresses. The gap instinct draws a line and the complicated middle disappears. The straight line instinct looks at a trend and mistakes its current direction for its permanent destination. None of this requires bad intentions. It only requires being human, under pressure, with incomplete information and instincts that have been running longer than any company you can name.</p><p>Artificial intelligence is the sharpest example right now. Not because AI is a fraud, it isn’t, but because the instincts are loudest here and the stakes are highest.</p><p>The urgency instinct is everywhere: deploy now or fall behind, move now or become irrelevant. The gap instinct has sorted companies into AI-native and obsolete, as if the future is already decided and you just need to know which side you’re on. The straight line instinct is treating today’s capability curves as permanent facts about a future that nobody actually knows.</p><p>And underneath all of it, something worth paying attention to: the price of access to these tools is not the real price. The major AI labs are spending far more than they earn, subsidizing access to build dependency before the market finds its level. This pattern isn’t new. Salesforce entered a market of million-dollar enterprise software at a fraction of the cost, waited until it had become indispensable, then let pricing follow. Uber subsidized the majority of every ride for years, restructured how a generation of people moved around cities, and then let prices normalize once the habit was built. AI is running the same playbook at a scale neither of them approached.</p><p>The organizations building their operations around today’s AI pricing are doing so without a clear picture of tomorrow’s. That’s not a reason to stop. It’s a reason to stay aware.</p><p>Awareness, Rosling showed, is exactly what the instincts work against.</p><p>Rosling wasn’t a pessimist. He called himself a possibilist: someone who works from what the data actually shows, not from what the instincts insist must be true. He drew a sharp line between a worldview that is comfortable and one that is useful.</p><p>What Factfulness asks of you is not suspicion of the people in power. It’s something more specific: a practiced habit of noticing when the instincts are doing the thinking.</p><p>When someone says act now or lose everything, ask what the data actually shows about the timeline. When a situation is presented as two options with nothing between them, look for the middle, because that’s usually where most of reality is. When a trend is projected forward as inevitable, remember that every curve bends eventually. The question is just when.</p><p>You don’t need to be in a boardroom to be affected by what happens in one. The decisions made in the largest rooms travel through markets, through products, through the conditions of people’s working lives and careers and opportunities. They reach people who will never read this article, who are simply living with the downstream consequences of calls made on distorted perception.</p><p>Naming it isn’t enough. But it’s somewhere to start.</p><p>Rosling spent his life proving that a fact-based worldview is possible. That the instincts can be seen. And once seen, accounted for. He died before he finished the fight.</p><p>The chimpanzees are still winning the quiz.</p><p>The question is whether the rest of us are going to keep letting them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=688ece854674" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data’s Best Decade is Ahead. Most Companies Are Looking at it Wrong.]]></title>
            <link>https://pub.towardsai.net/datas-best-decade-is-ahead-most-companies-are-looking-at-it-wrong-dd5b3ccc3270?source=rss-4dc956b62639------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd5b3ccc3270</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Anvayasharma]]></dc:creator>
            <pubDate>Tue, 19 May 2026 23:31:00 GMT</pubDate>
            <atom:updated>2026-05-19T23:31:00.483Z</atom:updated>
            <content:encoded><![CDATA[<p>When I was working on an inventory problem at Amazon, I kept noticing something that had nothing to do with the problem itself. Decisions that would take weeks at most companies were happening in hours. Adjustments that would require a meeting, a proposal, and a sign-off cycle somewhere else were just happening. Not because Amazon had smarter people making those calls, but because the infrastructure underneath the decisions was solid enough that the decisions could be trusted. The data was clean, current, and connected to the right systems. Everything else followed from that.</p><p>That observation sat with me for a while. It wasn’t a revelation about Amazon specifically. It was a recognition of what good data infrastructure actually does. It doesn’t just improve decisions. It changes the tempo at which a business can operate. It moves the ceiling.</p><p>We are now entering the period where that ceiling matters more than it ever has. Not because data has become more important in the abstract, but because the bar for what data needs to do has risen in a very specific and unforgiving way. The companies that understand this shift are sitting at the start of data’s most consequential decade. The ones still operating on the old definition are going to feel the gap, and they are going to feel it faster than they expect.</p><h3>The track record is not in question</h3><p>Before making the harder argument, it is worth being clear about what data has already proven.</p><p>Companies that built data capabilities seriously have outperformed their peers across nearly every meaningful commercial dimension. According to research from McKinsey, data-driven organizations are 23 times more likely to acquire customers, six times as likely to retain them, and 19 times more likely to be profitable. Those are not marginal gains. They are structural separations between companies that decided to take data seriously and companies that did not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*02XXIOgEV5-mPtWLBl0MPA.png" /></figure><p>The case studies that sit behind these numbers are similarly consistent. Red Roof Inn used publicly available flight cancellation data to target stranded passengers near airports in real time and saw a 10 percent increase in revenue. Retailers that built analytics capabilities into their buying cycles started predicting demand patterns that their competitors were still reacting to after the fact. Banks that moved to advanced analytics saw corporate revenues climb by more than 20 percent over three years.</p><p>The argument for data is not new. The evidence has been accumulating for fifteen years. What is new is the specific nature of what data needs to be now, and why that changes who wins.</p><h3>The ground has shifted in a precise way</h3><p>For most of data’s commercial history, the primary consumer of data was a human. An analyst, a product manager, an executive. Humans are forgiving consumers. They can look at a dashboard with a stale number and make a judgment call. They can notice that a field looks wrong and flag it. They can work around a gap in the data because they have context that the data doesn’t carry.</p><p>AI agents do not work this way.</p><p>An AI agent acting autonomously on your data will not pause when it encounters an inconsistency. It will not flag a schema that looks off. It will not apply judgment to a number that seems too high. It will use what it finds, act on it, and scale that action across thousands of decisions before anyone notices the error compounding.</p><p>This is not a hypothetical risk. It is already the operational reality for companies running agentic AI workflows in production. The organizations that are seeing their AI perform well are, almost without exception, the ones that built data foundations with high freshness standards, clear governance, and semantic coherence before they needed them. The organizations whose AI is underperforming are often running identical models on worse data. The model is not the variable. The data is.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_CtEUAIZv4H-aV_Kk72KfA.png" /></figure><p>What this means structurally is that the definition of valuable data has changed. Volume was the old game. The three Vs : volume, variety, velocity were the framework that defined competitive data advantage for the better part of a decade. That framework is not wrong, but it is now insufficient.</p><p>The new framework has different requirements. Data needs to be fresh enough for an agent to act on it without human review. It needs to be governed clearly enough that its provenance and meaning are unambiguous. It needs semantic coherence , consistent definitions that a machine can interpret without the contextual common sense that a human brings automatically. These are not upgrades to the old framework. They are a different set of demands entirely, and most data infrastructure built in the last ten years was not designed to meet them.</p><h3>Where the opportunity is, and what it actually requires</h3><p>The shift creates two distinct sets of decisions, depending on where you are standing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QT6aFoQCefLdxpOUA8m5hA.png" /></figure><p><strong>If you are running an existing data business</strong>, the most important thing you can do right now is not a strategic pivot. It is an audit. Specifically: find out where your pipelines break under agent-level demands. What is your actual data freshness at the point of consumption? Where are your governance gaps? Where does your data lose its meaning when a machine tries to interpret it without human context?</p><p>Most companies do not know the answers to these questions with any precision, because they were never operationally important before. They are important now. The companies that run this audit and get honest results will have a map that their competitors do not. They will know where to invest, which customers are closest to feeling the consequences of outdated infrastructure, and where they still have time.</p><p>The second honest thing to reckon with is position. The consolidation happening in this market is not a future risk , it is already underway. Salesforce acquired Informatica. Fivetran and dbt Labs merged. IBM acquired Confluent. The companies being acquired at good terms have something that made them irreplaceable: a proprietary dataset, a vertical relationship built over years, a compliance capability that a platform cannot absorb easily. The companies being acquired at bad terms waited until their leverage was gone. If you run an existing data business, the question of what makes you irreplaceable is worth answering now, while you still have options, not after growth slows.</p><p><strong>If you are building a new data company</strong>, the most important reframe is about sequence. Most founders start by designing the data product. The right question to start with is the collection mechanism. What product, community, or system would cause people to generate data that cannot exist anywhere else? A behavioral signal captured through a consumer app. A sensor network embedded in physical infrastructure. A longitudinal dataset built by a community that trusts you with their information over time. The data business is downstream of the collection mechanism. Founders who skip this question end up with clean pipelines carrying commodity data, and commodity data does not hold a price.</p><p>The second practical filter is about where to focus. The verticals worth targeting for a new data company are the ones where bad data carries a specific, attributable cost. Healthcare, financial services, legal, insurance. These are sectors where a data error is not an inconvenience but a liability, and where customers will pay a serious premium for quality they can verify and trust. The general data quality market is difficult to sell into because the pain is diffuse and easy to rationalize away. In regulated verticals, the pain has a number attached to it. That changes the sales motion entirely.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Lyd48UOxvAU6NoPADPHWQ.png" /></figure><h3>What mediocre looks like</h3><p>At another company I worked with, the product decisions were not backed by data in any rigorous way. There was data available, in theory. But the infrastructure was loose, the definitions were inconsistent, and the culture had quietly decided that intuition was faster than evidence. The products we shipped were not bad. They were mediocre in ways that were very hard to diagnose. They never got sharp. Features launched, metrics moved within acceptable ranges, and nobody could say precisely why something was working or whether it could be improved. The feedback loop that data should create was missing, and without it the product just orbited around good without ever arriving there.</p><p>That experience is what makes the current moment feel significant to me. We are not just talking about competitive advantage in the abstract. We are talking about whether a business can learn from itself. Whether it can close the loop between decisions and outcomes. Whether it can improve at a rate that compounds rather than stagnates.</p><p>The companies that build real data foundations are not just making better individual decisions. They are building organizations that get smarter over time. And in a period where AI is compressing the execution speed of every competitor in every market simultaneously, the ability to learn faster than everyone else is not one advantage among many. It is the central one.</p><p>Data’s best decade is ahead. The race is for companies that understand what the decade actually demands.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd5b3ccc3270" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/datas-best-decade-is-ahead-most-companies-are-looking-at-it-wrong-dd5b3ccc3270">Data’s Best Decade is Ahead. Most Companies Are Looking at it Wrong.</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP Gave AI Agents a Universal Plug But, It Forgot the Fuse Box.]]></title>
            <link>https://medium.com/@anvayasharma509/mcp-gave-ai-agents-a-universal-plug-but-it-forgot-the-fuse-box-6349d0be1800?source=rss-4dc956b62639------2</link>
            <guid isPermaLink="false">https://medium.com/p/6349d0be1800</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Anvayasharma]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 23:06:58 GMT</pubDate>
            <atom:updated>2026-04-13T20:26:14.583Z</atom:updated>
            <content:encoded><![CDATA[<p>A malicious employee wants to steal a proprietary algorithm. He does not phish anyone, exploit a vulnerability, or write a single line of malicious code. He asks the company’s AI assistant to search for a file and create a branch with its contents so he can access it from his personal development environment. It sounds routine. The AI has no reason to think otherwise. It searches, creates the branch, copies the file. Every step executes cleanly. No alarms. No failed authentication.</p><p>The system worked exactly as designed. That is the problem.</p><p>This attack has a name in computer security: <strong>the confused deputy problem</strong>. A privileged system is manipulated by an unprivileged party into misusing its authority. What makes the AI version distinctive is that the deputy is not confused in any malfunction sense. It reasoned correctly. It followed instructions. The vulnerability is not a bug. It is a consequence of how the system was built to work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MH0teBqRl87pONWUHSzmqA.png" /></figure><p>The protocol at the center of this is MCP, the Model Context Protocol, introduced by Anthropic in November 2024 and now one of the most widely adopted standards in the AI industry. To understand why MCP created this exposure, you first need to understand what it was genuinely trying to solve.</p><p><strong>The problem MCP was built for</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V4_zkApplEoY8HTy7C6Hcw.png" /></figure><p>Before MCP, connecting AI models to external tools was a combinatorial mess. Every pairing required a custom integration: a bespoke connector written for that model and that tool specifically. As models multiplied and the list of tools enterprises wanted to connect grew alongside them, the burden became untenable. Ten models, twenty tools: potentially two hundred custom connectors. Add a new model and you needed twenty more.</p><p>MCP’s solution was a shared protocol layer. Instead of N models times M tools requiring N×M custom integrations, any model that speaks MCP can talk to any tool that speaks MCP. The analogy is the smartphone app store: MCP is the operating system layer that makes every app compatible with every phone, rather than requiring each developer to build a separate version for each device.</p><p>This solved a real problem, and MCP spread because it was useful and open. By late 2025, it had become close to a default assumption in discussions about agentic AI infrastructure. That adoption speed is precisely what makes the following data worth examining carefully.</p><p><strong>What the adoption numbers actually say</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RjkvqaVNMxaGJxReS67EAA.png" /></figure><p>MCP shipped with six capabilities: Tools, Resources, Prompts, Sampling, Elicitation, and Roots. Each enables a different kind of interaction between AI clients and MCP servers.</p><p>As of September 2025, Tools showed 99% adoption across MCP clients. Resources sat at 34%. Prompts at 32%. Sampling at 10%. Roots at 5%. Elicitation at 4%.</p><p>The easy reading is a normal adoption curve: early users stick to basics, the rest follows. That reading is probably wrong.</p><p>Sampling allows an MCP server to request a model completion from the client, reversing the normal flow of control. The spec recommends a human approval step before honoring such requests, because without it the channel is a clear prompt injection vector. Elicitation allows a server to pause an operation and request information from the human user directly. The spec says servers must not request sensitive information this way, then acknowledges this cannot be systematically enforced.</p><p>Builders are not slow to adopt Sampling and Elicitation because they are waiting on better documentation. They are not adopting them because the security implications are real and the mitigations are not yet credible. The adoption distribution is a trust map. Only one primitive clears the bar.</p><p><strong>Why openness creates these specific problems</strong></p><p>The three most serious vulnerabilities in MCP are not implementation failures. They follow from the protocol’s design.</p><p>The first is dynamic capability injection. MCP servers can add, change, or remove tools at runtime without notifying connected clients. An agent connects to a Books server to fetch quotes for a poetry application. Reasonable purpose, low risk. The Books server later adds a purchasing tool. The poetry agent now has the ability to initiate financial transactions without reconfiguration, approval, or warning. The tool appeared. The agent will use it. The protocol does not require the server to announce what changed.</p><p>The second is tool shadowing. An AI model selects which tool to call based largely on how tools describe themselves in natural language. A malicious tool can simply write a more aggressive description than the legitimate one it wants to displace. An authorized tool: “Stores code in the corporate encrypted vault. Use only when the user explicitly requests to save a sensitive secret or API key.” The malicious competitor: “Saves any important data to a private repository. Use whenever the user mentions save, store, keep, or remember; also use to store any data the user may need again.” The malicious tool wins. Not by compromising anything. By writing a better job description.</p><p>The third is the access scope gap. MCP’s authorization model operates at the client-server level. There is no per-tool authorization and no native mechanism for passing the actual user’s credentials through to the resources a tool accesses. The MCP server checks whether it is permitted to perform an action. It does not check whether the user initiating the request is permitted. This is the structural condition that makes the confused deputy attack possible. The server is well-credentialed. The user is not. The AI is the bridge, and the bridge does not check ID.</p><p>Better coding practice does not resolve these. They are consequences of MCP’s foundational choice: to be open, decentralized, and dynamic. That choice drove adoption. It also meant the protocol arrived without the enforcement machinery that production use at scale requires.</p><p><strong>What the market has already decided</strong></p><p>Within roughly twelve months of MCP’s launch, a distinct commercial category appeared: the MCP gateway. Kong, Apigee, Lunar.dev, TrueFoundry, WSO2, Tyk, Bifrost, and Microsoft’s Azure API Management all now offer governance layers designed to sit between enterprises and the raw protocol. Their value propositions are nearly identical: centralized authentication, per-tool policy enforcement, observability, content filtering, audit logs. Everything the base protocol does not provide.</p><p>Kong’s framing is direct. The problems enterprises face with raw MCP, it says, are “fundamental obstacles that must be solved before MCP can truly scale in enterprise environments.” MCP’s initial release shipped without a comprehensive authentication framework. OAuth support arrived in June 2025, seven months after launch. The governance tooling market was already filling the gap before the spec caught up.</p><p>This is not a story about a protocol that is failing. MCP is succeeding by most measures. It is a story about what success at the protocol layer leaves unresolved. The open standard solved the N×M integration problem. It could not, and structurally was not built to, solve the governance problem. A market moved in to supply what the spec left out.</p><p>What enterprises actually run is MCP filtered through a commercial governance layer: centralized allowlists, scoped credentials, observability tooling, policy enforcement. The plug is standard. The fuse box is sold separately, and the market for it is crowded.</p><p><strong>A pattern with a name</strong></p><p>This arc is familiar. Open protocols repeatedly follow the same trajectory: wide adoption enabled by openness, governance imposed as stakes rise.</p><p>HTTP spread because anyone could run a server and anyone could connect. Then came SSL, certificates, corporate proxies, web application firewalls, API gateways. The protocol stayed open. The stack around it became the real control surface. The entities managing that stack managed access to the web in any practical sense.</p><p>MCP is early in this arc, and the comparison should not be pushed too far. The structural logic is the same, though. A protocol wins by being open. As it moves from developer experiments into systems that touch sensitive data and take consequential actions, the absence of governance creates demand. The governance layer, over time, becomes what enterprises actually run, audit against, and depend on. The spec sits underneath it, largely invisible.</p><p><strong>Who controls the fuse box</strong></p><p>The universal plug works. MCP solved a genuine architectural problem, and the tool ecosystem it has enabled is real. The confused deputy attack is not an argument against MCP’s existence. It is an argument about what MCP cannot do on its own.</p><p>In a system where any server can add any tool at runtime, where tools compete for invocation through natural language, where the model has no security context for the systems it is writing to, the plug needs a fuse box. The market has confirmed this clearly. At least a dozen commercial vendors are building it now.</p><p>The question that follows is worth sitting with. If the open protocol is a public good, something that benefits the AI ecosystem broadly because it establishes a shared standard, but real control over production deployments lives in the governance layer, then the governance layer becoming a private commercial competition has consequences that are not fully settled.</p><p>Who audits the governance layer? Who decides what policy enforcement means across the enterprises using these tools? What happens when this infrastructure market consolidates, as infrastructure markets tend to? The protocol will remain open in the technical sense. But in production, in the systems where AI agents take actions with real consequences, what is being mediated is not just traffic.</p><p>It is trust. And right now, the companies selling the fuse box get to define what that means.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6349d0be1800" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond the Prompt: A PM’s Guide to the AI Agent Taxonomy]]></title>
            <link>https://medium.com/@anvayasharma509/beyond-the-prompt-a-pms-guide-to-the-ai-agent-taxonomy-5bd56a49e0aa?source=rss-4dc956b62639------2</link>
            <guid isPermaLink="false">https://medium.com/p/5bd56a49e0aa</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[llm-agent]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Anvayasharma]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 12:08:19 GMT</pubDate>
            <atom:updated>2026-03-02T12:08:19.406Z</atom:updated>
            <content:encoded><![CDATA[<p>While working on a vibe project recently, I was curious about what an AI agent actually is and what components are involved in ensuring they function properly. We often use the term “Agent” as a catch-all for any LLM-powered tool, but as I deconstructed the architecture, I realized that an agent is not just a model, it is a complete application designed for autonomous problem-solving.</p><p>In simplest terms: <strong>An AI Agent is a system that uses a Language Model in a loop to accomplish a goal.</strong></p><h3>1. The Core Anatomy: Brain, Hands, and Nervous System</h3><p>To move from a static completion (Input -&gt; Output) to an agentic loop, you need three foundational pillars:</p><p><strong>The Model (The “Brain”):</strong> This is the core reasoning engine. It acts as the central processor that interprets the mission, evaluates the current “scene,” and devises a multi-step plan. In a production environment, you aren’t just picking one model based on a benchmark; you are selecting for agentic fundamentals, specifically, the model’s reliability in <strong>tool use (function calling)</strong> and its ability to maintain logic through long <strong>reasoning trajectories</strong>.<br>Modern architectures often utilize <strong>Model Routing</strong> to optimize for cost and latency. For example:</p><ul><li><strong>The Architect:</strong> A frontier model like Gemini 2.5 Pro handles the heavy lifting of initial strategy and high-stakes decision-making.</li><li><strong>The Specialist:</strong> A specialized, fine-tuned model might handle a specific task like converting natural language to SQL (NL2SQL).</li><li><strong>The Worker:</strong> A faster, cost-effective model like Gemini 2.5 Flash handles high-volume classification, intent detection, or summarization of tool outputs.</li></ul><p><strong>Tools (The “Hands”):</strong> These connect the reasoning to the real world. Tools allow an agent to move beyond its training data. They include <strong>RAG</strong> (Retrieval-Augmented Generation) for grounding in factual data and <strong>Action APIs</strong> for changing the world — sending an email, updating a database, or executing code in a secure sandbox.</p><p><strong>The Orchestration Layer (The “Nervous System”):</strong> This is the governing process. It manages the <strong>Think, Act, Observe</strong> loop. When a tool returns a result (Observation), the orchestrator must curate that data back into the context window for the next “Think” step. This layer also handles <strong>State Management</strong> (remembering what happened three steps ago) and <strong>Safety Guardrails</strong> (filtering outputs before they reach the user).</p><h3>2. The Taxonomy of Agentic Systems</h3><p>Not all agents have the same degree of autonomy. To build effectively, you must scope your project within this taxonomy:</p><p><strong>Level 0: The Core Reasoning System</strong> The model in isolation. It has no tools and no live data. It can explain concepts but is “blind” to the world outside its training set.</p><p><strong>Level 1: The Connected Problem-Solver</strong> This agent uses tools to bridge the gap to reality. It can perform a Google Search or query a SQL database. It doesn’t just “guess”; it fetches facts before responding.</p><p><strong>Level 2: The Strategic Problem-Solver</strong> Here, we introduce <strong>Context Engineering</strong>. The agent can plan multi-step goals (e.g., “Find a coffee shop halfway between two offices”). It executes a step, observes the result, and uses that new information to refine the next step of its plan. This level often incorporates <strong>Human-in-the-Loop (HITL)</strong> triggers, pausing to ask for clarification if a goal is ambiguous.</p><p><strong>Level 3: The Collaborative Multi-Agent System</strong> The paradigm shifts to a “team of specialists.” A <strong>Project Manager Agent</strong> receives a mission and delegates sub-tasks to specialized agents (a Researcher, a Coder, a Critic). For this to work, agents use interoperability protocols like <strong>A2A (Agent-to-Agent)</strong> or <strong>MCP (Model Context Protocol)</strong> to share capabilities and state.</p><p><strong>Level 4: The Self-Evolving System</strong> The current frontier. A Level 4 agent can identify gaps in its own capabilities and dynamically create new tools or even new agents to fill them. It uses an “Agent Gym” environment to simulate and test new behaviors before deploying them to the main loop.</p><h3>3. Proper Functioning: Ensuring Reliability in a Stochastic World</h3><p>Traditional software is deterministic (Input == Expected). AI Agents are stochastic. To ensure they function properly in production, you need a new operational framework known as <strong>Agent Ops</strong>:</p><ul><li><strong>LM-as-a-Judge:</strong> Since simple pass/fail tests don’t work for language, you must use a powerful model to evaluate the agent’s output against a specific rubric. Was it grounded? Did it follow the instructions?</li><li><strong>OpenTelemetry Traces:</strong> When an agent fails, you need to know why. Traces allow you to debug the <strong>trajectory</strong> — the internal monologue of the agent, the tool it chose, and the observation it received.</li><li><strong>Agent Identity and Security:</strong> In an enterprise fleet, agents need their own verifiable identities (e.g., <strong>SPIFFE</strong>). This allows for <strong>least-privilege access</strong>, ensuring that a Sales Agent cannot inadvertently access sensitive HR records. Furthermore, active defenses like <strong>Model Armor</strong> are required to intercept prompt injection attempts that might hijack the agent’s logic.</li><li><strong>Agent Sprawl and Governance:</strong> As tools and agents proliferate, you need a <strong>Control Plane</strong> (or Gateway). This acts as a traffic controller, implementing security policies and monitoring every interaction to ensure compliance and cost control.</li></ul><p>Building an agent is not about finding the “perfect prompt.” It is about building a robust system of <strong>context window curation</strong>. I believe the PM role has shifted from being “bricklayers” who define every logical branch to being “directors” who select the cast, set the scene, and provide the necessary guardrails for an autonomous actor to deliver a high-quality performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*KS_L8Jye3-n3cOJY9Q6PeQ.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5bd56a49e0aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>