Your contributors are AI-first now. Is your project?

The question of how to handle AI-generated contributions to a project has become a common discussion among maintainers. Nicholas Tindle, founding AI engineer at AutoGPT, shared his experience with this issue in a recent interview. AutoGPT, with over 180,000 stars and around 150 open pull requests, had a significant number of those PRs written by agents like Copilot, OpenClaw, and AutoGPT's own internal tools.

Initially, Tindle tried improving contributor guidelines and documentation to help agents, but these efforts didn't yield significant results. The problem lies in how agents interact with documentation - they don't read it unless explicitly instructed to. To address this, AutoGPT started placing instructions directly where agents look, such as CLAUDE.md files and AGENTS.md files that sit beside the code they govern. This placement allows skills to discover and load the necessary instructions dynamically.

Tindle found that centralizing the standard AGENTS.md and pointing Claude files at it helped resolve the issue. The placement of AGENTS.md beside the governed code mattered greatly, as it ensured that skills could access the necessary instructions, even if the agents couldn't locate them.

One useful nuance Tindle discovered was that AGENTS.md can be scoped to a directory, allowing skills to load dynamically and govern specific areas of the codebase. This approach enables front-end engineers to ship guides and skills tailored to their specific needs, like writing Storybook tests for components in specific folders.

To further improve the pull request process, AutoGPT implemented several gates that enforce certain rules, such as pull request templates, test plans, and Codecov coverage thresholds. These gates help ensure that pull requests meet specific criteria before they can be merged, reducing the need for manual intervention and improving the overall quality of contributions.

One key takeaway from Tindle's experience is the importance of requiring a CLA (Contributor License Agreement) to detect human contributors and require a commit SHA before resolving a review thread. By doing so, the team can ensure that human oversight is maintained while still allowing AI-generated contributions to be processed efficiently.

Finally, AutoGPT learned that the CI (Continuous Integration) process should be treated as a strict requirement rather than a suggestion. By enforcing coverage thresholds and ensuring that pull requests meet specific criteria before merging, the team can maintain high-quality standards without relying on manual checks.