<?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 Roy Yu on Medium]]></title>
        <description><![CDATA[Stories by Roy Yu on Medium]]></description>
        <link>https://medium.com/@iroy2000?source=rss-3e3fbc385d03------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*fDGj07XyKcHKUIVB.</url>
            <title>Stories by Roy Yu on Medium</title>
            <link>https://medium.com/@iroy2000?source=rss-3e3fbc385d03------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 14:45:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@iroy2000/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[I Tried to Automate My Own Job. Here’s What I Found.]]></title>
            <link>https://medium.com/@iroy2000/i-tried-to-automate-my-own-job-heres-what-i-found-15fb86d415c2?source=rss-3e3fbc385d03------2</link>
            <guid isPermaLink="false">https://medium.com/p/15fb86d415c2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Roy Yu]]></dc:creator>
            <pubDate>Tue, 31 Mar 2026 15:16:01 GMT</pubDate>
            <atom:updated>2026-04-06T15:34:35.744Z</atom:updated>
            <content:encoded><![CDATA[<p>There’s a conversation happening inside almost every engineering team right now.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/727/1*LC8V1Pl0HRuKH6WtsCUkjw.png" /><figcaption>The moment execution becomes automated, intent becomes everything.</figcaption></figure><blockquote><strong>Is AI going to take my job?</strong></blockquote><p>Most of us don’t say it out loud.<br>It feels too vulnerable. Too career-limiting. Almost embarrassing, as if asking the question means you already know the answer.</p><p>So instead, we post about <em>“AI as a co-pilot”</em> and <em>“humans in the loop”</em>, and then go back to our daily routine.</p><blockquote><strong>I got tired of pretending.</strong></blockquote><p>So I ran an experiment.</p><h3>The Setup</h3><p>I’m a software engineer.</p><p>One of the typical day-to-day work looks like this:<br>Pick up a ticket &gt; Understand the codebase &gt; Write code &gt; Write tests &gt; Open a PR &gt; Respond to comments &gt; Iterate &gt; Merge &gt; Repeat.</p><p>So I built two things.</p><p><strong>The first:</strong> a virtual developer.<br> Give it a ticket, and it reads the requirements, explores the codebase, writes the implementation, adds tests, runs checks, opens a pull request, complete with a summary comment back on the ticket, and waiting to respond to PR comments.</p><p>End-to-end. Unattended.</p><p>But I don’t want to stop here, I want to bring up my point of “got tired of pretending”. What if it can do it in scale.</p><p><strong>The second:</strong> an orchestrator.<br>Instead of one ticket, give it many. It figures out dependencies, sequences what needs to be serialized, and spins up parallel agents, and each working in its own branch.</p><p>The ultimate goal here is</p><blockquote><strong>While I slept, it will work through my backlog</strong></blockquote><p>Real tickets.<br>Real codebase.<br>Real pull requests.</p><h3>What Actually Happened</h3><p>Some of it was genuinely breathtaking.</p><p>Tasks that would’ve taken me an afternoon — API integrations, config migrations, utility functions — were done in minutes.</p><p>Clean PRs.<br>Tests passed.<br>Commit messages that made sense.</p><p>If you didn’t know better, you’d assume a decent engineer wrote them on a good day.</p><p>The orchestrator handled dependencies correctly.<br>Avoided merge conflicts.<br>Posted ticket updates.</p><p>It even responded to code review comments — occasionally pushing back, in my own voice, when a reviewer was wrong.</p><p>For about twenty minutes, I felt something close to vertigo.</p><p>I was watching something do my job.</p><h3>The Elephant in the Room</h3><p>Let’s name it directly.</p><p>If a tool can pick up a ticket, implement it, write tests, open a PR, and respond to review…</p><p><strong>What exactly is the engineer doing?</strong></p><p>A meaningful portion of the job most of us were hired to do… just ran on a laptop while I made coffee.</p><p>Something has changed.</p><p>The question is what.</p><h3>Garbage In, Garbage Out</h3><p>Here’s what I actually learned:</p><p>The output quality was almost perfectly correlated with the input quality.</p><p>Not my code. That’s just artifacts of what AI produced.</p><p><strong>But the source of truth, in my case, my tickets.</strong></p><p>When a ticket was clear — well-scoped, with solid acceptance criteria and context — the agent produced work I’d be proud to ship.</p><p>When a ticket said something like:</p><blockquote><strong><em>“Investigate session handling issue”</em></strong></blockquote><p>…the agent produced confident-sounding nonsense.</p><p>And if I’m honest, that describes at least a third of my backlog.</p><p>The same pattern showed up in code review.</p><p>Precise, grounded comments → correct fixes.<br>Vague feedback like <em>“this doesn’t feel right”</em> → guesswork.</p><p>Sometimes right. Sometimes very wrong.</p><blockquote><strong>The AI didn’t break things.</strong></blockquote><blockquote><strong>The humans did.</strong></blockquote><p>Not because they’re bad engineers — but because the inputs were unclear, the requirements ambiguous, the feedback imprecise.</p><p>The AI was just a mirror.</p><p>How much of my day is spent compensating for unclear requirements?<br>How many “technical decisions” are just educated guesses?<br>How often is my feedback precise enough that someone else could execute it faithfully?</p><h3>What the Loop Is Actually For</h3><p>I used to think <em>“humans in the loop”</em> (<strong>HIL</strong>) meant catching AI mistakes.</p><p>And yes — there were mistakes.</p><p>Missed edge cases.<br>Questionable abstractions.<br>Locally correct decisions that didn’t fit the bigger picture ( btw, this happened a lot to me ).</p><p>But that wasn’t the real role of the human.</p><p><strong>The real role was intent.</strong></p><p>There were moments where:</p><ul><li>The agent implemented exactly what the ticket asked for — and the ticket was wrong.</li><li>A review comment was addressed literally — but missed the spirit.</li><li>A decision was logically correct — but contextually wrong.</li></ul><p>The system worked.</p><p>The instructions didn’t.</p><p>The loop doesn’t exist to catch hallucinations.<br>It exists because software development is full of tacit knowledge — things no one writes down.</p><p>And right now, a human is still the one carrying that.</p><h3>What This Actually Means</h3><p>I don’t think AI is going to wipe out engineering jobs overnight.</p><p>But it <em>will</em> hollow out a certain kind of work:</p><p><strong>Mechanical execution of well-specified tasks.</strong></p><p>If your value is writing code against a clean spec — that’s at risk.</p><p>If your value is:</p><ul><li>defining what the spec <em>should</em> be</li><li>navigating ambiguity with product and design</li><li>making architectural tradeoffs ( brainstorm with AI )</li><li>writing feedback precise enough to act on</li></ul><p>— that’s much harder to replace.</p><p>This experiment made something uncomfortably clear:</p><p><strong>The bottleneck was never writing code.</strong></p><p>It was deciding what to write.</p><p>That’s not a coding problem.</p><p>It’s a judgment problem.</p><blockquote><strong>And judgment comes from understanding — <br>from context, from experience, from caring about the outcome.</strong></blockquote><h3>The Fear, Revisited</h3><p>I started this experiment half-hoping it would scare me.</p><p>But that’s not what I found.</p><p>What I found was this:</p><p>The thing most worth protecting isn’t my ability to write code.</p><p>It’s my ability to:</p><ul><li>think clearly about problems</li><li>communicate precisely</li><li>hold a coherent point of view over time</li></ul><h3>The Real Shift</h3><p>The elephant in the room isn’t that AI will replace engineers.</p><p>It’s that AI will make something visible — very quickly:</p><blockquote><strong>Which engineers were mostly executing…<br>and which ones were actually thinking.</strong></blockquote><p>The uncomfortable part?</p><p>You don’t get to hide behind the process anymore.</p><p>The good part?</p><p>You still get to choose which one you become.</p><p><em>I built the tooling described here on top of Claude. Since I’m using my company laptop to do this experiment, I can’t share the codebase, but I’m happy to share the approaches and techniques if there’s interest.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15fb86d415c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Automation Didn’t Remove Responsibility. It Made It Invisible.]]></title>
            <link>https://medium.com/@iroy2000/automation-didnt-remove-responsibility-it-made-it-invisible-96a3ecd3e1e6?source=rss-3e3fbc385d03------2</link>
            <guid isPermaLink="false">https://medium.com/p/96a3ecd3e1e6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[society]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Roy Yu]]></dc:creator>
            <pubDate>Mon, 09 Feb 2026 17:00:09 GMT</pubDate>
            <atom:updated>2026-02-09T17:39:15.977Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><em>Automation made decisions faster, but not more neutral.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fuIRsai_R6JCr225xO4goQ.png" /></figure><p>Something essential has quietly disappeared.</p><p><strong>Visibility.</strong></p><p>“Human in the loop” used to imply oversight and accountability. But now that involvement is invisible.</p><p>When decisions affect someone’s livelihood or reputation, people are rarely told whether a human reviewed them.</p><p><strong>Automation didn’t eliminate responsibility…</strong></p><p><strong>It obscured it.</strong></p><p>Trust doesn’t collapse because systems are imperfect, it collapses when accountability is unclear.</p><p>Visibility is not anti-AI.</p><p>We disclose ingredients, calories, and security practices because people deserve to know.</p><p>In the AI era, disclosure of human involvement should be no different.</p><p>“Closed by a Human” is a public signal initiative. It means a real person reviewed and approved a decision.</p><p>Not that it will make things perfect. But it is about visibility and that someone is accountable.</p><p><strong>Progress without accountability isn’t innovation.</strong></p><p><strong>It’s abdication.</strong></p><p>More here: <a href="https://closedbyahuman.com/">https://ClosedByAHuman.com/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96a3ecd3e1e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>