For 18 years GitHub has treated pull requests (PRs) as open by default. Now several AI-first open-source projects are reversing that practice: they automatically close external PRs and rely on automated agents to triage, reproduce, implement and review contributions. Project teams say their own agents and a ‘‘software factory’’ model reduce backlog and speed up review cycles.
The software factory concept and Vercel’s AI SDK
Vercel described building a ‘‘software factory’’ for its open-source AI SDK, which receives more than 20 million npm downloads per week. By late June the project had a backlog of “over 1,000 open issues and almost 800 pull requests,” so the team deployed agents to regain control of the backlog.
Vercel’s system uses different types of agents for distinct tasks: one reproduces bugs, another applies fixes, and another performs reviews. Lars Grammel, a Vercel engineer, told a YouTube audience that the team trusts its own, carefully optimized agent prompts more than community-run agents. He argued that a well-tuned agent configuration that has historically succeeded at fixing a category of bugs builds trust and can reduce review time.
Grammel outlined the deployment architecture: a UI, a web app, an underlying API, an execution space and sandboxes, all synchronized with GitHub to trigger automated actions. Vercel reports that four weeks after implementing the software factory, the system ‘‘authors between 25 and 35% of PRs we merge and closes 70–80% of issues.’'
Astro’s auto-triage and organizational shift
The Astro web framework, which has 62,000 stars on GitHub, has adopted the software factory idea under creator Fred Schott. Schott told Latent Space that for five years issues arrived faster than the team could handle; in the last six months, agent-based triage has fundamentally changed that.
Astro’s automation reproduces bugs, proposes fixes, and asks users to verify bot-suggested fixes before human maintainers intervene. Schott said the workflow has shifted how the team handles incoming requests — issues are now items they prioritize weekly rather than a continuously trimmed backlog. The auto-triage work on Astro directly led Schott to build a new agent framework called Flue.
Flue and tldraw: converting PRs into issues and discussions
Flue’s contributor guide states the project will ‘‘reimagine things’’ and aims to prevent what it calls “Drive-by AI slop PRs.” In Flue, every external pull request is automatically closed and converted into an issue or a discussion: bug reports and fix proposals become issues, feature requests become discussions. The team combines its expertise with “the best available SOTA LLMs” to decide what to prioritize, and then deploys agents for research, design, implementation and initial review.
Similarly, tldraw (50,000 stars) announced in January that it would automatically close external PRs; project founder Steve Ruiz reiterated the decision five months later. Ruiz framed the policy as a response to changing coding practices (more discussion, more agents), evolving social practices around public contribution, and shifting risks around code security.
Mitchell Hashimoto, HashiCorp co-founder and Ghostty creator, now a co-founder of Superlogical, suggested a further possibility: that large open-source projects may ultimately close contributions entirely.
Implications for community contributors
Traditionally, maintainers reviewed PRs not only to check code but also to teach contributors and evaluate them as prospective maintainers. If projects increasingly use agents for code review and implementation, it raises questions about opportunities for community members to take on more active development roles.
Fred Schott acknowledges the gap: narrowing contribution channels risks problems if core maintainers become unavailable. However, both Flue and tldraw still accept new issues and discussions, which encourages dialogue. That conversational route can help community members get to know and trust one another, learn from peers, and potentially demonstrate they deserve maintainer privileges.
Steve Ruiz summarized the practical trade-off: if it is easier for maintainers to have their agents write code than to accept external patches, community contributions may be best focused where they remain most valuable — reporting issues, participating in discussions, offering perspective and care.
Conclusion
A number of AI-native open-source projects have moved away from accepting external PRs in favor of internal agent-driven workflows and ‘‘software factory’’ architectures. Teams report measurable reductions in backlog and faster resolution, but the approach changes how communities participate and how contributors become maintainers. Projects are currently trying to balance efficiency gains with preserving conversational and reporting channels that allow community members to demonstrate commitment and build trust.
