Vercel's seven-agent "software factory" now writes a third of merged PRs on the AI SDK

Open issues on the AI SDK went from 1,022 to 844 in four weeks. That's the number I'd lead with if I were writing this post, and Vercel did.

Vercel published the internals of the system it built to maintain the AI SDK, a repo with over 20 million npm downloads a week, 26,000+ GitHub stars, and 100+ new issues a month. Four weeks in, the factory authors 25-35% of merged PRs weekly, closes 70-80% of incoming issues, accounted for more than 75% of July's closures, and dropped open bugs about 25%. The stack is boring on purpose: Vercel Functions, Queues, Sandbox, Neon Postgres.

The architecture is the part worth copying. There are seven agents and none of them is "the maintainer agent." They're narrow: classification, analysis, implementation, automated review, bug reproduction, bug fixes, backporting. Each one has a job small enough that you can write a pass/fail check for it. Bug reproduction is separate from bug fixing, which is separate from review. That decomposition is why the thing works at all, and it's the opposite of what most people build when they wire up their first agent pipeline. The instinct is one big agent with fifteen tools. What actually ships is a pipeline of small agents with three tools each.

The second detail: nothing merges without human approval. At a 20M-download-a-week blast radius, Vercel put the gate at merge and nowhere else. Agents triage, reproduce, patch, and review each other, and then a person clicks. That's the correct gate placement, and it maps to what Anthropic's runtime-safety position paper argues this same week (arXiv 2608.11274): enforcement belongs in the harness, not in a system prompt asking nicely.

What I'd do with this: pick the single most mechanical part of your maintenance loop and give it one agent. For most repos that's issue classification, because it's cheap, high-volume, and wrong answers are recoverable. Don't start with bug fixes. Vercel's numbers are real but they come from a repo with dense test coverage and an unusually legible issue stream. Your repo probably doesn't have that, and the agent will happily produce confident garbage in the gaps.

The uncomfortable part is the follow-on question nobody at Vercel answered: what happens to the junior maintainer role when the factory closes 75% of issues? Herrengt's HN post this week (912 points, 815 comments) argues implementation-to-spec is exactly what's being commodified. I don't think he's right about the timeline, but I think he's pointing at the right layer.