On August 12, GitHub published the AutoGPT maintainer playbook. Nicholas Tindle, AutoGPT's founding AI engineer, shared how a project with 180,000 stars and roughly 150 open PRs—many of them generated by agents—stays maintainable now that agent-submitted code is the norm.
This is not a tutorial on writing AGENTS.md, but a manual for redesigning open-source repositories when agent-generated PRs have become standard practice.
Key points
AutoGPT's current state
- 180,000 stars
- Roughly 150 open PRs, a large portion written by agents
- AGENTS.md is the third version of AutoGPT's instruction files—previous versions were CLAUDE.md files, ignored by Copilot and Codex because they don't recognize Claude's private filename
- A Storybook-testing guide for frontend, triggered when components are in certain directories
- The same rule enforced on the backend via a coverage threshold: 80% or no PR opens
- PR template wording triggers a test-PR skill where the agent installs agent browser, starts the app, and runs the changes
- Old thinking: educate agents to write better PRs → govern agent behavior
- New thinking: turn repository rules into the only entrance agents can pass → make governance part of the product itself
- Uneven agent PR quality
- Reviewer fatigue
- Inconsistent rule enforcement
- Backlog accumulation
- Multi-agent collaboration (Anthropic's research the same day addresses this separately)
- Safety of agent-submitted code in long-tail scenarios
- Attribution—who is responsible for ultimately merged code
Tindle's core framing: "It's basically somebody else paying for your compute." If a contributor wants to spend tokens improving your project, let them—but the contract must be written into the repository.
Five composable guardrails
1. One standard instruction file, where agents can see it. AutoGPT standardized on AGENTS.md; all scattered CLAUDE.md files point to it. AGENTS.md sits in the same directory as the code it governs, because agents read the files in front of them, not wiki links.
2. Skills with trigger phrases. A skill is an instruction file with a description telling agents when to load it. Examples built into the repo:
The result: the team almost never receives PRs that fail to run.
3. PR templates as a behavioral wall. Non-conforming PRs are auto-closed without hesitation. Tindle wrote the automation first but found it unnecessary—agents already comply with the template before the bot runs. The rule changed behavior; enforcement never had to fire.
4. CLA as a human detector. AutoGPT's CLA requires a browser + GitHub OAuth flow + a separate domain—a combination agents handle poorly. PRs unsigned after one week are closed and reopened after re-signing.
Tindle recommends every project do this, including top-tier MIT-licensed ones.
5. Commit SHA as a precondition for resolving reviews. Some agents mark all review threads resolved without changing code. AutoGPT's pr-address skill defines the only legal sequence:
> fix → commit → push → reply → resolve
The reply must link the full SHA of the fix commit, obtained via git rev-parse HEAD. The skill explicitly lists anti-patterns: "Acknowledged" is not a fix, and citing a commit that doesn't touch the flagged lines isn't either.
What was shut down
The first iteration, a "CI-failure commenting agent" wiring Claude Code into GitHub Actions, effectively added a broadly scoped credential inside CI. Tindle turned it off—anyone who could run CI could then access all of Claude Code's capabilities with the same credential.
The real significance
Many teams' first response to agent-submitted PRs is "find more reviewers" or "write stricter rules." The playbook inverts this: don't try to convince agents to behave—restructure the repository so non-compliance can't get in.
What it solves—and what it doesn't
Solved:
Not solved:
Quick-start checklist for other projects
1. Write a real AGENTS.md—not vibes: concrete rules (allowed paths, coverage thresholds, PR size, commit format)
2. Link AGENTS.md from CONTRIBUTING.md so both humans and agents can see it
3. Add a PR template + auto-close bot
4. Introduce a machine-parseable rules block (JSON) that GitHub Actions can read directly
The whole setup takes about 1 hour—most of it spent on judgment calls like "which directories really need CODEOWNERS gating," not on configuration.
The real value: when 90% of PRs come from agents, a repository's maintainability depends not on agent politeness, but on the repository's own contract design.