Building a software factory for AI SDK
Vercel’s AI SDK factory shows a practical scaling pattern: narrow agents produce evidence inside sandboxes while humans retain merge authority and review effort follows risk.
AI SDK had accumulated **over 1,000 issues and almost 800 pull requests** by late June. Four weeks after introducing its factory, Vercel says agents authored **25–35% of merged PRs** and closed **70–80% of issues**, while humans approved every merge.
The reusable pattern is one agent per reviewable task, each with its own prompt, context, and evals. Start locally, pass evidence between classification, analysis, reproduction, implementation, and review stages, then vary human scrutiny by change risk rather than treating every agent output equally.
AI SDK had accumulated **over 1,000 issues and almost 800 pull requests** by late June. Four weeks after introducing its factory, Vercel says agents authored **25–35% of merged PRs** and closed **70–80% of issues**, while humans approved every merge. The reusable pattern is one agent per reviewable task, each with its own prompt, context, and evals. Start locally, pass evidence between classification, analysis, reproduction, implementation, and review stages, then vary human scrutiny by change risk rather than treating every agent output equally. These are early results from one large open-source project, not a controlled comparison. The factory also depends on isolated sandboxes, restricted secrets and networking, queues, monitoring, and sustained human review, so the headline automation rates omit substantial operating machinery.
This turns prior harness guidance into an early operating result: a staged, evidence-passing factory handled a substantial share of one project’s issue and PR flow while retaining human merge authority. It reinforces reviewable task boundaries, isolation, and risk-based approval, but does not show that the reported rates transfer beyond AI SDK or survive without significant orchestration and monitoring.