Top AI-native open source projects replace community pull requests with agent-driven software factories

Vercel’s open source AI SDK, which receives over 20 million npm downloads weekly, deployed a software factory composed of specialized agents to regain control over a backlog that had ballooned to more than 1,000 open issues and nearly 800 pull requests by late June. The system includes agents for reproducing bugs, applying fixes, and reviewing changes before handing results to humans for merging. Within four weeks of deployment, Vercel reports the factory now authors between 25% and 35% of the pull requests it merges and closes 70% to 80% of issues.

Astro’s web framework, with 62,000 GitHub stars, adopted a similar software factory approach to triage and resolve issues. Creator Fred Schott said the team had been in a state where issues arrived faster than maintainers could address them, but automation has since restored balance. The system now handles triage, reproduction, and even suggests fixes that users can verify before maintainers review them.

Flue, a new agent framework created by Schott, takes this model further by automatically closing external pull requests and converting them into issues or discussions. The project’s contributor guide frames this as a way to prevent “drive-by AI slop PRs” and to reframe contributions as leads rather than code submissions. Once issues or discussions are prioritized, Flue’s agents perform research, design, implementation, and initial review before handing work back to humans.

tldraw, a source-available React drawing tool with 50,000 GitHub stars, also automatically closes external pull requests. Project creator Steve Ruiz announced the policy in January and reiterated it five months later, citing changes in coding practices, social norms around public contribution, and code security as reasons for the shift. The project converts external PRs into issues or discussions and uses agents to implement changes when appropriate.

Maintainers describe greater trust in their own agentic systems than in community-generated code. Vercel engineer Lars Grammel explained that when a narrowly scoped agent with a specific prompt has a strong historical track record, the team develops confidence in that configuration. “For open-source projects, it’s worth considering having your own agents and your own setup, and not necessarily trusting the community,” Grammel said, noting that this approach can reduce review time.

The shift raises questions about community participation and maintainer succession. Schott acknowledged the risk that narrowing contribution pathways could leave gaps when maintainers are unavailable, but he suggested that increased discussion and peer interaction could help community members prove themselves and learn from one another. Ruiz similarly argued that limiting code contributions to reporting, discussion, and perspective allows maintainers to focus on areas where human input still adds unique value.