Contributors are now AI-first. Are projects?

AutoGPT has over 180,000 GitHub stars and approximately 150 open pull requests, a significant number of which were written by AI agents such as Copilot, OpenClaw, and in-house tools. When maintainers sought to block agent-generated PRs, AutoGPT chose to allow agents to use computational resources to improve the project while clearly defining pass criteria.

The issue was not the volume of documentation, but how agents discovered it. Since agents prioritize instructions located directly before the working directory over wikis or guides for the entire repository, AutoGPT placed AGENTS.md near the code and configured files for Claude to point to common guidelines.

It also utilized skills that are dynamically loaded beyond directory scope. For example, modifying a component in the frontend folder automatically triggers a skill that guides the creation of Storybook tests, while the backend enforces an 80% test coverage condition.

Key quality gates include:

• Enforcing PR templates: PRs that do not follow the template are automatically closed, with notifications to adjust agent behavior proactively.
• Linking test plans: Specific phrases in the template invoke the test PR skill to install a browser agent and run the app.
• Operating CI as a mandatory gate: Codecov coverage thresholds are set as required checks, prompting failed agents to load the test skill.
• Using CLA as a human verification step: The signing process requires a browser on a separate domain and GitHub OAuth, mandating human intervention.

After implementing these measures, AutoGPT received more PRs that actually worked but did not align with the project roadmap, rather than PRs that failed to run. From a maintainer's perspective, controlling quality and direction through rules and automation within the repository creates a better problem than blocking agent productivity.