Leading AI-native open source projects like Flue and tldraw are rejecting external pull requests, partly because they are often AI-generated. Instead, maintainers use their own agents to manage contributions. Vercel's AI SDK project deployed a 'software factory' with specialized agents to handle its PR and issue backlog, improving efficiency and reducing review time.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
As AI-generated PRs flood in, top open source projects are closing external PRs and using their own agent teams to handle contributions. This isn't just an efficiency revolution—it may change the very nature of open source communities.
GitHub's PR mechanism has been open for 18 years, but now some top AI-native open source projects are closing PRs because they've found a better way. Projects like Flue and tldraw refuse to accept external contributors' PRs, partly because these PRs are often AI-generated. Maintainers prefer to use their own agents to create and manage PRs.
tldraw founder Steve Ruiz announced this policy in January and reaffirmed it five months later, citing changes in coding practices (more discussion, more agents), the social practice of public contribution, and the changing landscape of code security. Flue's contribution guidelines explicitly state that every external PR is automatically closed and converted into an issue or discussion to prevent 'Drive-by AI slop PRs.'
Vercel's AI SDK project has over 20 million npm downloads per week, but by the end of June it had accumulated over 1,000 open issues and nearly 800 PRs. To address this, Vercel deployed multiple agents responsible for reproducing bugs, applying fixes, and reviewing fixes. Engineer Lars Grammel said they trust their own optimized agent configurations over community agents because it shortens review time.
Just four weeks after implementing the software factory, Vercel claimed that the factory wrote 25-35% of merged PRs and closed 70-80% of issues. The Astro framework (62,000 stars on GitHub) has adopted a similar philosophy. Founder Fred Schott said that for the past five years, issues have flooded in faster than they could handle, but with agent-based triage, they've regained control.
Schott said: 'In the last six months, things have completely changed. We can now use these automations to handle these issues—triage, reproduce, have users verify bot-suggested fixes, and then we look at them.' The result is not only a significant reduction in open issues but also a complete change in how they are handled, something he says he's never seen in a decade of open source experience.
Traditionally, PR review is not just code checking but also a way to teach and evaluate future maintainers. If AI SDK and Astro use agents for most review and implementation, how can community members participate more actively? Schott acknowledges this risk: 'If you keep narrowing the scope of the project, what happens if you and I go on vacation? This doesn't solve everything.'
But Flue and tldraw, while not accepting PRs, do accept issues and discussions, which might point to a solution: through more communication, community members get to know and trust each other, learn from peers, and potentially prove themselves worthy of becoming maintainers. Ruiz suggests: 'It's best to limit community contributions to where they still matter: reporting, discussing, providing perspective, and showing care.'
This article is based on a report from Latent Space. Vercel's data (25-35% PRs, 70-80% issues) comes from Vercel's own claims and has not been independently verified; quotes from Astro and Flue come from founders and are source claims.
AI-native open source projects are shifting from 'open PRs' to 'agent factories,' significantly improving efficiency, but the model of community participation is being reshaped, potentially focusing more on discussion than code contribution in the future.
Primary report
Primary source