<?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 Coley Brown on Medium]]></title>
        <description><![CDATA[Stories by Coley Brown on Medium]]></description>
        <link>https://medium.com/@colebrow.cb?source=rss-17239080d8d8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*RYx5RracfuMvHk81.</url>
            <title>Stories by Coley Brown on Medium</title>
            <link>https://medium.com/@colebrow.cb?source=rss-17239080d8d8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:52:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@colebrow.cb/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[Knowledge Debt is the New Technical Debt]]></title>
            <link>https://medium.com/@colebrow.cb/knowledge-debt-is-the-new-technical-debt-22d16a3e9a4d?source=rss-17239080d8d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/22d16a3e9a4d</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[web-development]]></category>
            <dc:creator><![CDATA[Coley Brown]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 17:00:07 GMT</pubDate>
            <atom:updated>2026-10-27T17:00:07.754Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BQlYLpZ0NTtr-n-X" /><figcaption>Photo by <a href="https://unsplash.com/@officestock?utm_source=medium&amp;utm_medium=referral">Sebastian Herrmann</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I want to start with a question. Think of the most recent feature you built from scratch with AI. Could you diagram it right now, from memory? The data flow, the new pieces, how they connect to what was already there?</p><p>I hope the answer is yes. For a lot of us, if we’re being real, it’s no. As vibe coding and AI-assisted development move into enterprise teams, one skill is slipping away: holding a mental model of the system in your head. That skill is what separates engineers who can answer questions from engineers who have to go look things up.</p><h4>A year ago</h4><p>A year ago, I prided myself on being able to answer almost any technical question about our stack without digging through the code, That wasn’t because I’m special. It was because my hands were in every PR and technical discussion.</p><p>We’d walk through a ticket in grooming and I’d already be sketching a plan in my head. Bigger initiatives got a dedicated technical grooming where we went deeper. The code was written by hand, with some help from Copilot autocomplete. It was slow, but every line I wrote or reviewed added to my picture of where the product was headed.</p><h4>Today</h4><p>Fast-forward to today, and I feel less involved than ever. We still run the same agile ceremonies. The difference is that AI now does most of the building, and the understanding that used to come along with writing the code doesn’t come with it anymore.</p><p>By most measures, we’re doing great. We’re shipping more features than ever. But I feel left behind, and I don’t think I’m the only one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DlNzEclMoyb2whoC" /><figcaption>Photo by <a href="https://unsplash.com/@mlenny?utm_source=medium&amp;utm_medium=referral">Alexander Hafemann</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>What knowledge debt is</h4><p>Every engineer knows technical debt. You can take a shortcut to ship faster, and you pay for it later with interest. Knowledge debt works the same way, except the shortcut isn’t in the code. It’s in our heads.</p><ul><li><strong>The principal </strong>is every part of the system that nobody on the team really understands. It’s code that works and that we approved, but couldn’t explain without asking the AI to explain it back to us.</li><li><strong>The interest</strong> is slower debugging, longer incident calls, nervous refactors, and onboarding where the answer to every question is “ask the AI”.</li><li><strong>It comes due </strong>at the worst possible time: a 2am outage, a key engineer leaving, or a feature that has to change in a way the original design never planned for.</li></ul><p>There’s one big difference from technical debt. You can see technical debt. It shows up in the code as TODOs, workarounds, forked repos, and that one file everyone avoids. Knowledge debt is invisible. The code can look clean while the team’s understanding of it sits at zero, and you don’t find out until something breaks.</p><h4>How AI piles it on</h4><p>The problem we keep running into is that AI tries to build the perfect solution. Not the perfect solution for our project, our architecture, and our team.</p><p>When I started a new feature, my first instinct was always to reuse the tools and patterns we already had. That made the work easier to understand. PRs were smaller, and they were built from concepts the team already knew. Each one added a little to what we all understood.</p><p>AI doesn’t have that instinct unless we give it one. Hand it a ticket and it will happily bring in a new abstraction, a new library, or a new way of doing something we already do three other ways. The result is bigger, more complicated chunks of work. Each one might be fine on it’s own, but each one adds to the debt, because it’s something new the team has to learn after it’s already been merged.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LAULsOj9AhkYkYEE" /><figcaption>Photo by <a href="https://unsplash.com/@imlst?utm_source=medium&amp;utm_medium=referral">Valery Fedotov</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>“But the AI can just explain it”</h4><p>The obvious pushback is: who cares if you understand it? You can ask the AI anything about the codebase and get an answer in seconds.</p><p>That’s true, and it’s useful. But looking something up isn’t the same as understanding it. AI can answer the questions you know how to ask. Understanding is what tells you which questions to ask in the first place. It’s what lets you look at a proposal and say “that’s going to fight with how we handle caching” before anyone writes a line. You can’t spot a bad design in a system you’ve never really taken in.</p><p>That judgement is the most important part of my job. It’s the part I feel slipping.</p><p><strong>What we’re changing: plan without AI</strong></p><p>Our team is about to try one big change. <strong>We’re going to plan without AI</strong>, and then hand the AI a finished plan instead of a ticket. The idea is simple but it’s a fundamental change to our workflow.</p><p>Here’s how it works today. A ticket with a short description and half-written acceptance criteria goes straight to an AI agent. The AI fills every gap with it’s own assumptions, and we find out what it decided when we review the PR. By then the design decisions are already made, and none of them are ours.</p><p>The plan is to move the thinking back to the team, in the room:</p><ol><li><strong>We design it together. </strong>In technical grooming, we’ll work out how we’d build the feature ourselves: which existing services and patterns to reuse, where the new code lives, what the data flow looks like, and what we’re deliberately <em>not</em> doing.</li><li><strong>We write the plan down.</strong> That means complete acceptance criteria, the components involved, the patterns to follow, and the edge cases we already know about. If we can’t write it down, we don’t understand it well enough to built it yet.</li><li><strong>AI gets the plan, not the ticket.</strong> The agent will start with the context it’s been missing: our architecture, our conventions, and the decisions we’ve already made.</li><li><strong>We review against the plan.</strong> A PR review stops being “does this work?” and becomes “does this match what we agreed upon?” When the output goes off-plan, it should be easy to see.</li></ol><p>AI will still write most of the code, and we’ll still move fast. The difference is that the mental model gets built where it should, in the team’s heads, before the code exists. The goal is to pay down knowledge debt up front instead of letting it pile up after the merge.</p><p>We’re also betting the output will get better. Most of the overbuilt, “perfect in general” solutions come from AI guessing at context we never gave it. A complete plan should remove most of the guesswork.</p><p>We should easily be able to gauge our success with smaller PRs, fewer rewrites, and faster reviews.</p><h4>Back to the diagram</h4><p>AI isn’t the problem. Letting it make our design decisions for us is. Technical debt taught us that speed now always gets paid for later. Knowledge debt is the same lesson, just harder to see. There’s an old saying: “If you want to go fast, go alone. If you want to go far, go together.” AI lets us go incredibly fast. Planning as a team is how we make sure we go far.</p><p>So go back to the question I started with. Pick the last feature you built with AI and try to diagram it. If you can’t, that’s your knowledge debt. The fix isn’t to stop using AI. It’s to make sure the thinking happens before the prompt.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22d16a3e9a4d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agile Was Built for Human Bandwidth. AI Just Removed the Limit.]]></title>
            <link>https://medium.com/@colebrow.cb/agile-was-built-for-human-bandwidth-ai-just-removed-the-limit-d0a117384e1b?source=rss-17239080d8d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0a117384e1b</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Coley Brown]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 22:17:08 GMT</pubDate>
            <atom:updated>2026-09-29T22:17:08.167Z</atom:updated>
            <content:encoded><![CDATA[<p>For years, most of the software industry has relied on Agile development to ship code, and I’ve drunk the Kool-Aid more than once. The pitch is simple: break work into small, digestible chunks, iterate quickly, and deliver more often.</p><p>It worked because of a basic truth about people: we’re single-threaded. We do our best work on one problem at a time, and every context switch costs us. The talented engineers who could juggle multiple projects had developed an unusual skill. Agile didn’t try to change that. It built a process around it: one developer, one ticket, one sprint at a time.</p><p>That design made sense for a long time. The problem is that the developer is no longer the only one doing the work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ioa1FqhfHtdhefQY" /><figcaption>Photo by <a href="https://unsplash.com/@sent1nel4s?utm_source=medium&amp;utm_medium=referral">GABRIEL CARVALHO</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>The executor changed</h4><p>It’s like being handed a modern car when every road is still built for horse and buggy. The car is faster, but the roads won’t let you use it.</p><p>Agentic development changed who executes the work. I can have an agent building a large feature in the background while I investigate a performance problem, and another one performing a version upgrade. Yet most of our planning still assumes each engineer works on one ticket at a time.</p><p>To be clear, Agile’s values aren’t the problem. Responding to change, shipping working software, and tight feedback loops matter more now than ever. What’s breaking is the cadence and the ceremonies, all built around a unit of capacity that no longer exists: one human, one task, one sprint.</p><h4>What we know now</h4><p>If we started from scratch and built a methodology for modern development, three observations would shape it.</p><p><strong>Code is cheap.</strong> I’m not diminishing the work. I’m recalibrating what “output” means. The value is less about the code itself and more about whether it satisfies the requirements. This has been a contentious topic on my team: how do you make a pull request useful when you didn’t write the code? The syntax is no longer under a microscope, so review becomes an architectural discussion. The goal isn’t to pit models against each other in search of a perfect solution. It’s to maximize output while managing risk.</p><p><strong>Work can happen in parallel.</strong> We have been given a superpower but we are still figuring out how to harness it. Being assigned one ticket at a time already feels like a bygone era.</p><p><strong>Groomed work is the new bottleneck. </strong>If you had an employee who could work 24 hours a day, 7 days a week, and you still handed them tasks on a 9–5 schedule, you’d be severely underutilizing them. Agents execute faster than we can define what they should build. The scarce resource isn’t typing time anymore. It’s clear, well-scoped requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x5PF4YDMrmg0G4j7" /><figcaption>Photo by <a href="https://unsplash.com/@tashakostyuk?utm_source=medium&amp;utm_medium=referral">Tasha Kostyuk</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>Context switching didn’t disappear. It moved.</h4><p>This part rarely makes the hype posts. Running three agents doesn’t free you from context switching. It shifts the switching from writing code to reviewing it. Instead of losing focus mid-implementation, you’re jumping among three streams of output, each needing judgement. Any new process has to account for that load, or we’ll burn people out while calling it productivity.</p><h4>What a post-Agile process might look like</h4><p>We don’t have it figured out, but here’s the shape emerging on our team:</p><ul><li><strong>Plan around agentic capacity, not developer hours. </strong>Sprint planning built on story points per person no longer maps to what a team can ship.</li><li><strong>Make grooming the core engineering skill. </strong>If defined work is the bottleneck, writing clear specs and acceptance criteria is the highest-leverage thing an engineer can do.</li><li><strong>Treat review as a design discussion. </strong>Line-by-line review gives way to questions about architecture, risk, and whether the solution fits the system.</li><li><strong>Delegate ongoing maintenance to scheduled agents. </strong>We’ve started using scheduled tasks for reoccurring work. Every day, an agent can check the most common production error and propose a fix. That’s asynchronous work that doesn’t need a human until it’s time to approve a change.</li></ul><h4>The part nobody wants to talk about</h4><p>Many people go into engineering for the satisfaction of solving hard problems. Historically, that meant iterating on code until it was nearly perfect. AI has taken over most of that work, and it’s hard to find a meaningful replacement.</p><p>For me, it was never really about the code. It was about building something from nothing. And I’ve never shipped more quality features than I do now, yet I feel like I deserve less and less of the credit.</p><p>That may be the hardest problem of all. Any methodology we build next has to answer two questions, not one: how do we get the most out of these tools, and how do we make sure the people using them still feel like they built something?</p><p>Our team is still searching for that balance. I suspect most organizations are as well. One thing can be certain though, we are having a lot of fun figuring it out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d0a117384e1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ ClawdBot — With Great Power Comes Great Responsibility]]></title>
            <link>https://medium.com/@colebrow.cb/clawdbot-with-great-power-comes-great-responsibility-5e188261253d?source=rss-17239080d8d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e188261253d</guid>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[project-management]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Coley Brown]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 18:27:04 GMT</pubDate>
            <atom:updated>2026-02-02T18:27:04.387Z</atom:updated>
            <content:encoded><![CDATA[<h3>🦀 ClawdBot — With Great Power Comes Great Responsibility</h3><p>Let’s start with ClawdBot’s intended purpose. The promise on its website is simple and ambitious: <em>“The AI that actually does things.”</em> While I believe that promise is directionally true, ClawdBot is not the magical, hands-off solution the internet sometimes makes it out to be.</p><p>For starters, this is a new and emerging tool. That is evident from the fact that its name has already changed twice since I began experimenting with it: ClawdBot, then MoltBot, and now OpenClaw. For the sake of clarity, and habit, I will continue referring to it as ClawdBot throughout this article. Much like how I still call it Twitter instead of X.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mIS5XAFs_BOe8Nd6" /><figcaption>Photo by <a href="https://unsplash.com/@boliviainteligente?utm_source=medium&amp;utm_medium=referral">BoliviaInteligente</a> on <a href="https://unsplash.com?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>ClawdBot is a developer tool first and foremost. Out of the box, its functionality is fairly limited. Its greatest strength is also its greatest weakness: the ability for agents to take direct control of your machine with broad access to your browser and file system. For every article praising its capabilities, there are several others highlighting serious security concerns, and those concerns are understandable.</p><p>Despite being only weeks old at the time of writing, ClawdBot already has strong community support. That is clearly demonstrated by the extensive list of third-party skills available on moltbotskill.com, many of which significantly extend what the tool can do. After spending many hours working through my own local setup, I wanted to share what I learned and why I believe tools like this represent an important step toward the future of human-AI collaboration.</p><h4>Agents Need Structure, Not Just Power</h4><p>One of the biggest takeaways for me, and seemingly for others, is how we interact with agents going forward. Many posts showcase ClawdBot being used to generate a Kanban board for task management and progress tracking. At first glance, that might sound trivial. In practice, I think it points to something much more important.</p><p>The software industry has spent decades refining how development teams work together. You can move from one company to another and still find familiar processes. Agile development is a core example. It emphasizes quick iteration, tight feedback loops, and over-communication. These practices exist because they work.</p><p>When I started using ClawdBot, the first task I gave my primary agent, Jarvis, was to build a tool that mirrored how I already work. That tool ended up being a Kanban board. The goal was simple: I wanted visibility into what my AI agents were doing and how tasks were progressing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8VKfiacz8kqiv2NNgpZX2Q.png" /><figcaption>Mission Control, which was completely generated by ClawdBot, became the central interface for managing my agents and their work.</figcaption></figure><h4>Breaking Down Agent Responsibilities</h4><p>This setup becomes significantly more effective as your agent team grows in both capability and responsibility. In my case, I split functionality across four distinct agents, each optimized for a specific role.</p><p>While a single, powerful agent like Jarvis, running on the Opus 4.5 model, could technically perform all tasks, doing so would be inefficient and unnecessarily expensive. Instead, I separated concerns so each agent handles the work it is best suited for.</p><p>Here is the agent structure I, and many others, have converged on:</p><p><strong>Jarvis</strong></p><ul><li>Role: Main orchestrator</li><li>Responsibilities: Coordination, planning, reviews</li><li>Model: Opus 4.5</li></ul><p><strong>Coder</strong></p><ul><li>Role: Developer</li><li>Responsibilities: Code changes, debugging, feature development</li><li>Model: GPT-5 Codex</li></ul><p><strong>Assistant</strong></p><ul><li>Role: General support</li><li>Responsibilities: Documentation, organization, routine tasks</li><li>Model: GPT-4.1</li></ul><p><strong>Researcher</strong></p><ul><li>Role: Research and analysis</li><li>Responsibilities: Comparisons, investigations, reports</li><li>Model: GPT-4o</li></ul><p>To keep costs manageable, Jarvis acts primarily as a delegator. Think of it as a project manager with the highest level of context and reasoning. Tasks created on the Kanban board can be assigned to a specific agent, which automatically begins work once the task moves into an “in progress” state.</p><p>This structure closely mirrors how a high-functioning development team operates and allows me to maintain a familiar mental model while working with AI. While LLMs are powerful, they are ultimately constrained by unstructured text and subjective interpretation.</p><h4>Cost, ROI, and Security Tradeoffs</h4><p>Cost is one of the most obvious drawbacks of a tool as flexible as ClawdBot. During the initial setup and while building the Mission Control app, Jarvis burned through my ten dollars of Claude credits surprisingly quickly. That experience made one thing clear: cost visibility is not optional.</p><p>The next feature I prioritized was tracking model usage, cost estimation, and overall metrics. Ultimately, any tool is only as useful as its return on investment. If I am spending a hundred dollars to mark emails as read or build throwaway prototypes, I have to seriously question what value I am getting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wuqHBxXBu3ekK6wx0MZJCg.png" /><figcaption>Model Metrics and Cost Estimation</figcaption></figure><p>As I grew more confident in my setup, another question kept coming up: what <em>should</em> I automate? Adding more capabilities to agents also means increasing risk. Granting autonomous agents full browser and file system access raises serious security concerns. Power without guardrails is not something to take lightly.</p><h4>A Glimpse of What’s Coming</h4><p>Working with ClawdBot feels a lot like using the iPhone when the App Store first launched. No one knew exactly what kinds of applications would emerge, but everyone could sense the potential. That same feeling exists here.</p><p>ClawdBot is not production-ready, and that is okay. What matters is the direction it points in. It offers a glimpse into a future where AI systems do more than respond to prompts. They collaborate, execute, and operate within structured workflows.</p><p>I can easily imagine a future where tools like this are built directly into operating systems like macOS, complete with refined user interfaces and strong security safeguards. That is when we may truly see the promise of AI that actually does things fully realized.</p><p>Until then, I will keep experimenting with ClawdBot and similar tools, staying cautious, curious, and optimistic about where this technology is heading.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e188261253d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>