When PRs Stop Being Welcome: The Software Factory With Agents

For 18 years the pull request was open source's social contract: anyone opens a PR, a maintainer reviews it, discusses, teaches, and merges. A report by Richard MacManus on Latent Space documents a reversal of that contract among AI-native projects: Vercel (AI SDK), Astro, Flue, and tldraw are closing external PRs by default and delegating the work to teams of agents. The stated reason is direct: much of the contributions arriving today are generated by AI without care, and maintainers trust their own calibrated agents more than third-party code.

For the Brazilian developer who maintains (or dreams of maintaining) an open source AI project, the question isn't ideological. It's operational: how do you scale code review when the volume of contributions grows faster than your capacity to read diffs.

What Is a "Software Factory"

The term used in the report describes a pipeline where a team of agents, not a single copilot, handles a contribution's entire lifecycle. Each agent has a narrow function:

• a triage agent classifies the issue (bug, feature, noise);
• a reproduce agent reproduces the bug in an isolated sandbox;
• an implement agent implements the fix or feature;
• a review agent reviews what was implemented;
• a human does the final merge.

The key lies in specialization. Lars Grammel, an engineer at Vercel, explains in the source why the company prefers its own agents over the community's:

If we have a very specific agent with a very specific prompt that we've optimized, and we know that, over time, it has been very successful at fixing a certain category of bugs, then we develop trust in that agent configuration.

>

-- Lars Grammel, engineer at Vercel

This is the central technical argument: trust is a property of configuration, not of a person. An agent with a versioned prompt, measurable history, and a defined scope is auditable in a way that an anonymous PR from an unknown contributor never will be. Grammel describes the AI SDK's system architecture as a custom UI, a web app, an API, an execution space, and sandboxes, all synced with GitHub, which automatically triggers the next actions.

Vercel's Numbers

The AI SDK has surpassed 20 million weekly downloads on npm. By the end of June, according to the source, the backlog had exceeded a thousand open issues and nearly 800 pull requests. Four weeks after deploying the factory, Vercel claims it:

| Metric | Result (4 weeks) | |---|---| | Merged PRs authored by the factory | 25% to 35% | | Issues closed by the factory | 70% to 80% |

A dose of skepticism is warranted: these numbers are self-reported by Vercel, without external audit, and "closing an issue" is not synonymous with "solving a problem" (closing due to duplication or being unreproducible also counts). Still, the order of magnitude shows that the real bottleneck of modern open source isn't writing code, it's triaging and reproducing what comes in.

Astro, a framework with 62,000 stars on GitHub, tells a similar story. Creator Fred Schott describes five years in which "issues arrived faster than we could handle." With agents handling triage, reproduction, and even having the user verify the bot-suggested fix before any human looks at it, he says he has managed, for the first time in more than a decade of open source, to treat issues as something that can be prioritized weekly instead of an endless backlog.

The Radical Case: PRs Automatically Become Issues

Astro's auto-triage system gave rise to Flue, an agent framework by Schott himself that takes the policy to the extreme. In Flue, every external PR is automatically closed and converted: a bug report or fix proposal becomes an issue; a feature request becomes a discussion. The project's contribution guide uses the term "drive-by AI slop PRs" for what it wants to avoid.

The metaphor Schott uses is interesting for anyone thinking about product: treat incoming contributions as leads, not as work the maintainer feels obligated to review. The decision of what to build next comes from the issue or discussion, combining the team's expertise with "the best SOTA LLMs we have access to." Only after the decision is made do agents step in for research, design, implementation, and initial review.

tldraw (50,000 stars), a drawing tool built in React under a "source available" license, adopted the same automatic PR closure. Creator Steve Ruiz described the change as "an opinionated decision in response to changes in how we're coding (more discussion, more agents), the social practices around public contribution, and the shifting code security landscape." Mitchell Hashimoto, HashiCorp cofounder and creator of Ghostty, goes further in the source and bets that "the future is that large open source projects will close contributions entirely."

What Breaks: The Pipeline of New Maintainers

Here's the trade-off the report itself doesn't hide. Historically, reviewing a PR was never just about the code. It was how a maintainer taught a contributor and evaluated them as a future maintainer. If agents do the review and implementation, the natural mechanism for developing people disappears.

Schott himself acknowledges the risk: "there's still that gap of: well, if you just keep narrowing the project down, at some point you and I are going to go on a trip, and then what happens? This doesn't solve every problem."

The hint of a solution the source suggests is that Flue and tldraw close PRs but open issues and discussions. The idea is that the community gets to know each other and build trust by talking, and that's where someone proves they deserve to become a maintainer, rather than proving it via diff. Ruiz sums it up as better to limit community contribution to the places where it still matters: "reporting, discussion, perspective, and care."

What This Changes for Builders in Brazil

For the Brazilian maintainer, the model doesn't require Vercel's custom infrastructure to start paying off. Here's what you can extract without a proprietary UI or expensive SOTA models:

• Separate triage from implementation. The biggest reported gain (70% to 80% of issues closed) came from triage, not from writing code. A GitHub Actions workflow that runs an LLM to classify issues and request minimal reproduction already tackles the main bottleneck.
• Force reproduction before a human gets involved. Astro's practice of having the issue author verify the bot-suggested fix before any human looks at it is cheap and filters out a lot of noise.
• Version and measure your agent prompts. The trust Grammel talks about doesn't come out of nowhere: it comes from track record. Without measuring success rate by bug category, you don't know which agent to trust.
• Don't copy full PR closure too early. Projects like Flue and tldraw have a large base and a paid team. A small Brazilian project that closes PRs might just be scaring away the few contributors it has, without mature agents to make up for it. Closure makes sense when the volume of "slop" already costs more than what human contribution delivers.

The point that remains open, and that the source honestly leaves unanswered, is human sustainability: an ecosystem where code comes from agents and new maintainers no longer have the PR as an entry point needs to invent another door. Discussion and issues are a start, but they're not yet a proven path for developing people. For those maintaining AI projects in Brazil, it's worth following this less as a trend to copy and more as an ongoing experiment, gathering your own numbers before closing any door.