PRs NOT Welcome: How Top AI Open Source Projects Are Managing Thousands of Contributors
Latent Space Richard MacManus
Some top AI open source projects are shutting off outside PRs. They’d rather use their own agents than review AI slop from strangers.
Based on reporting by Latent Space, Richard MacManus — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
GitHub made pull requests the default way to contribute, and that model has held for 18 years. Now a small but prominent slice of AI-native open source is backing away from it. Projects like Flue and tldraw are refusing external PRs altogether, in part because so many of them are now AI-generated. Their maintainers would rather route work through agents they control themselves.
Vercel’s AI SDK is the clearest example of where this is headed. In a late-June backlog, the project had more than 1,000 open issues and almost 800 pull requests. Vercel says its newly built “software factory” now uses separate agents to reproduce bugs, apply fixes, review the result, and hand it back to a human only for the merge. Four weeks after the system went live, Vercel said it was authoring 25 to 35% of merged PRs and closing 70 to 80% of issues.
That shift is not just about volume. Vercel engineer Lars Grammel said the company trusts its own tightly tuned agents more than agents run by community members, because it knows how those agents behave over time. He also laid out the setup in practical terms: a UI, a web app, an API, an execution space, and sandboxes, all synced with GitHub. The UI was custom-built.
Astro has taken a similar route. Fred Schott said the project spent five years buried under issues faster than it could handle them. Over the last six months, though, agents have taken over triage, reproduction, and even getting users to verify a bot’s suggested fix before the maintainers step in. Astro says that has not just reduced the backlog; it has changed the team’s whole rhythm around incoming community requests.
Schott liked the pattern enough to build Flue, where external PRs are automatically closed and converted into issues or discussions. tldraw does the same. Steve Ruiz says that if the issue is already well specified and agents can write the code, outside code contributions matter less than reporting, discussion, perspective, and care. That is a pretty blunt way of saying the old open source ritual is being replaced by something more managed, and maybe more efficient.
My take — AI-written commentary, not fact-checked reporting
The awkward truth is that a lot of open source contribution was always only partly about code. It was also about filtering people, teaching people, and building a queue of future maintainers. The new agent-heavy model gets the work done, but it risks turning communities into suggestion boxes with nicer branding. That’s efficient. It’s also a little bleak, which is very on-brand for the current state of software.
Read more about this at: Latent Space
