The same question keeps coming up in maintainer conversations: what do you do when the pull request queue fills with work written by agents?
It’s something Nicholas Tindle, founding AI engineer at AutoGPT, also deals with every day.
I spoke with him in May for Maintainer Month.
At the time of the interview, AutoGPT had over 180,000 stars and around 150 open pull requests.
A big chunk of those pull requests were written by agents, including Copilot, OpenClaw, and AutoGPT’s own internal tooling, among others.
Most maintainers I talk to have the same reaction: close the door.
Turn off pull requests.
Don’t tax the team with reviewing slop.
Nicholas saw an upside: It’s basically somebody else paying for your compute.
Nicholas Tindle, founding AI engineer at AutoGPT The way he sees it, if a contributor wants to spend their tokens improving your project, let them.
Just make it so the only way through the door is the way that works for you.
Your docs aren’t the problem.
Discovery is.
AutoGPT tried the obvious thing first.
Better contributor guidelines.
Better docs.
A whole wiki dedicated to working with the repo.
None of it moved the needle.
It turns out the tools aren’t going to go read your docs unless they’re told to.
That’s the part a lot of us get wrong.
We treat documentation like the agent will go find it.
It won’t.
Agents read what’s in front of them, at the level of the directory they’re working in.
So AutoGPT started putting instructions where agents look.
First files, because laude was generating pull requests without enough repository-specific context.
The commit trailer made each one easy to spot, because they announced themselves in the commit trailer.
Then they hit the next wall: Copilot and Codex ignore Claude files, because they’re not Claude.
So they centralized the standard and pointed Claude files at it.
Here’s the nuance I found most useful. is scoped to a directory.
A skill can be discovered outside that directory. (If you haven’t shipped one: a skill is an instruction file with a description that tells the agent when to load it.
The agent scans descriptions up front and pulls in the full instructions when the task matches.) AutoGPT’s sits beside the code it governs.
That placement matters as much as the instructions themselves.
If you’re writing backend tests and you think about doing front-end stuff, a skill may load dynamically.
It’s not going to know what directory to go look in for an AGENTS.md file, but the skill can tell it that.
Their front-end engineer got tired of the same class of broken pull request, so they wrote a guide, and shipped it as a skill in the repo.
The description contained trigger phrasing: write a Storybook test if your component lives in these folders.
Now every harness that touches the repo discovers it automatically.
The backend enforces its own version of the rule the same way: hit 80% coverage or don’t open the pull request.
Gates that actually work These are the gates you can adapt for your project.
Enforce the pull request template, loudly.
AutoGPT tells agents that pull requests not matching the template get closed automatically with zero hesitation.
They built the tooling to actually do it, then found they didn’t need to run it.
At AutoGPT, the rule changed agent behavior before the automation ever ran.
The agents followed the template.
Human contributors sometimes needed more room, which Nicholas treats as a feature: If you don’t follow the template, I know you’re probably a person, and I’m going to be kinder.
The test plan trick.
The template requires a test plan, and its wording casually mentions testing the pull request.
That phrase triggers a skill called , which installs agent browser (with permission), spins up the app, and executes the change.
The agent set out to fill in a checkbox and ended up running the code.
They almost never get pull