<?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 Tomas Achmedovas on Medium]]></title>
        <description><![CDATA[Stories by Tomas Achmedovas on Medium]]></description>
        <link>https://medium.com/@minvs1?source=rss-f80030097fb1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ApdkGhMJ-Y2sbTjHiHrxlw.jpeg</url>
            <title>Stories by Tomas Achmedovas on Medium</title>
            <link>https://medium.com/@minvs1?source=rss-f80030097fb1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 09:25:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@minvs1/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[Two agent loops that quietly handle our backlog]]></title>
            <link>https://medium.com/@minvs1/two-agent-loops-that-quietly-handle-our-backlog-c82348ee1a46?source=rss-f80030097fb1------2</link>
            <guid isPermaLink="false">https://medium.com/p/c82348ee1a46</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Tomas Achmedovas]]></dc:creator>
            <pubDate>Fri, 19 Jun 2026 13:18:28 GMT</pubDate>
            <atom:updated>2026-06-19T13:18:28.108Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C9JdXDTHl1iLyibYyN-kLw.jpeg" /></figure><p>This is a small setup we run at trip1, but it takes a real share of our routine tickets off us. It’s two loops and two clicks. One loop watches our Linear backlog and reliably flags the tickets an agent can handle on its own. The other takes a ticket once we’ve approved it, opens a PR, and runs it between a writer agent and a reviewer agent until it’s clean. From our side it’s two clicks: a label to approve a ticket, and the merge at the end. We don’t really prompt the agents anymore. We wrote the prompts into the loops once and stopped touching them, and the clicks are the only part we still do by hand.</p><h3>The first loop</h3><p>The first loop is a Claude Code Routine on a schedule. It reads the backlog and decides, per issue, whether an agent could finish it without a human: clear scope, contained change, nothing in a sensitive part of the code. It labels the ones that qualify with a one-line reason and ignores everything else.</p><p>Then a person adds an approval label to the candidates they want built. It’s the first of the two clicks, and it isn’t a safety check. Triage flags more than we want open as PRs at once, and a stack of PRs nobody can review is useless, so the label is how we meter how many candidates become actual work. I expect we’ll drop it once triage is reliable enough to let candidates open PRs directly and handle the volume some other way.</p><h3>The second loop</h3><p>The second loop is another Routine. We didn’t want to run a glue service to fire it off the label, so it runs on a schedule too and picks up whatever’s been approved since the last pass. It reads the ticket, makes a branch, writes the change and tests, runs the suite and linter, and opens a PR linked to the issue. From there a GitHub Action we already had reviews the diff and tags each comment blocking or minor. The agent fixes the blocking ones and pushes, the Action runs again on the new commits, and that repeats until nothing blocking is left and CI is green, capped so it can’t loop forever. Then it asks for a human review.</p><p>That review is the second click. Branch protection requires a human approval, so the agents can’t merge their own work. We’re starting to let the lowest-risk changes merge automatically.</p><h3>What we constrain</h3><p>What we constrain is reach. Size doesn’t matter much; which part of the codebase a change touches does. Each agent gets the minimum access it needs, and the build agent can only push to branches with a set prefix, so it can’t reach anything protected. There’s a list of code it never touches at all: auth, payments, secrets, infrastructure, migrations. A ticket that needs any of those goes to a person. A big refactor somewhere boring is fine. A one-liner in the payment path isn’t.</p><p>The whole thing runs on the Claude subscription, not the API. That was a hard requirement. Something that runs on a schedule and reacts to every push shouldn’t bill per token.</p><h3>trip1</h3><p>trip1 is the hotel booking site we run. You book a stay and pay in crypto, across a few million properties. It also supports x402, the agent-payments protocol, so an agent with its own wallet can book a room and pay in USDC on Base without a human or a card. The same works through our MCP server, so an agent can search and book a hotel end to end. If you’re building agents that need to actually buy something in the world, it’s at <a href="https://trip1.com/">trip1.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c82348ee1a46" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>