Your contributors are AI-first now. Is your project?
GitHub highlights AutoGPT’s approach to managing AI-generated pull requests, emphasizing structured guidelines and automated gates to maintain code quality while accommodating AI contributors.
Video
Video available
AutoGPT, with over 180,000 stars and 150 open pull requests, faced a growing challenge: many contributions were generated by AI agents like Copilot and OpenClaw. Maintainers often responded by restricting access, but AutoGPT took a different approach. Nicholas Tindle, founding AI engineer, explained that instead of blocking AI contributors, the project adapted by placing instructions where agents naturally look, such as CLAUDE.md and AGENTS.md files. This shift aimed to guide AI tools without requiring them to seek out documentation independently.
The project refined its strategy by centralizing instructions in AGENTS.md files, scoped to specific directories. Skills—instruction files with descriptions—were introduced to dynamically load relevant guidelines. For example, a frontend engineer created a skill triggering Storybook tests for components in designated folders. AutoGPT also enforced pull request templates, which agents followed automatically, while human contributors received more flexibility. This approach reduced broken pull requests and improved code quality without additional manual intervention.
AutoGPT implemented automated testing as part of the pull request process. The template required a test plan, which triggered a skill to install a browser, spin up the application, and execute the change. This ensured that pull requests were tested before review, drastically reducing non-functional submissions. While most pull requests now worked, some still didn’t align with the project’s roadmap—a preferable issue compared to broken code. Continuous integration (CI) was treated as a strict requirement, with tools like Codecov enforcing coverage thresholds automatically.
To further filter contributions, AutoGPT used a Contributor License Agreement (CLA) as a human detection mechanism. Signing the CLA required a browser-based GitHub OAuth flow, a step most AI agents cannot complete. Unsigned pull requests were closed after a week, reintroducing human oversight. Additional gates included requiring commit SHAs to resolve review threads and using Copilot in workflows to comment on failed checks. These measures ensured that AI contributions met the project’s standards while minimizing manual maintenance burdens.