Closing the door on agent-written PRs solves the wrong problem

The maintainer reflex when the pull request queue fills with agent-written work is to close the door: turn off pull requests and spare the team from reviewing slop ^[5]. Nicholas Tindle, founding AI engineer at AutoGPT, treats that reflex as solving the wrong problem, because the queue is not going away and the contributors filling it have no human author to coach ^[2].

When GitHub's Andrea spoke with Tindle in May for Maintainer Month, AutoGPT had over 180,000 stars and around 150 open pull requests, a large share of them written by agents including Copilot, OpenClaw, and AutoGPT's own tooling ^[1][2][3][4]. Tindle's framing is that a contributor spending tokens on your project is "basically somebody else paying for your compute" ^[6]. The job is not to block that work but to make the only way through the door the way that works for you.

AutoGPT tried the obvious remedy first, better contributor guidelines, better docs, a dedicated wiki, and none of it moved the needle ^[7]. The reason is mechanical: agents do not go and find your documentation, they read what is in front of them at the level of the directory they are working in ^[8]. So the team moved instructions to where agents actually look. They started with CLAUDE.md files, because Claude was opening pull requests without enough repository context, and the commit trailer made each one easy to spot ^[9]. Then Copilot and Codex ignored those files because they are not Claude, so AutoGPT centralized the standard AGENTS.md and pointed the Claude files at it ^[10].

The distinction they found useful: AGENTS.md is scoped to a directory and sits beside the code it governs, while a skill, an instruction file with a description that tells an agent when to load it, can be discovered from outside that directory ^[11]. The front-end engineer, tired of the same broken pull request, shipped a guide as a skill with trigger phrasing about writing Storybook tests for components in certain folders; the backend enforces its own rule, 80 percent coverage or do not open the pull request ^[12].

The gates are built for agents but read humans as the exception. AutoGPT tells agents that pull requests not matching the template are closed automatically, and it built the tooling to do it, then found the rule changed agent behavior before the automation ever ran ^[13]. Human contributors sometimes needed more room, which Tindle treats as a signal rather than a failure: "If you don't follow the template, I know you're probably a person, and I'm going to be kinder" ^[14]. The template also requires a test plan, and its wording casually mentions testing the pull request, a phrase that triggers a skill called test PR that installs an agent browser with permission, spins up the app, and runs the change, so the agent sets out to fill a checkbox and ends up executing the code ^[15].

What to watch is whether this inverts the usual review economics. The bet is that when instructions live where agents read and gates enforce themselves, most of the queue conforms before a human looks, leaving scarce review time for the contributors who break the rules because they are people.