Building a software factory for AI SDK
Not a pilot. Vercel's software factory authored 25-35% of AI SDK's weekly PRs in four weeks—and closed 70-80% of issues without removing humans from control.
Why it matters
A production-grade agent system handling real open-source infrastructure work demonstrates how autonomous systems scale maintenance when humans stay in the loop. The architecture—sandboxed agents, per-task specialization, risk-aligned review depth—is a blueprint for agentic engineering at scale.
The key facts
11 to knowAI SDK receives 100+ issues/month; backlog hit 1,022 open issues + 800 PRs by June 2026
Factory agents authored 25-35% of merged PRs within 4 weeks of production launch
70-80% of issues closed by factory; open issues fell from 1,022 to 844 in 6 weeks
Factory PRs comprise 50%+ of v6 backport merges; v5 backports previously skipped due to merge conflict burden
July 2026: 75% of closed issues were factory-authored
Open bugs down ~25% since factory launch
Architecture: dedicated agents for classification, analysis, implementation, review, backporting, docs; each sandboxed with task-specific secrets and network shielding
Built incrementally starting with issue classification; now 6 specialized agents in production pipeline
Every change requires human approval before merge; factory provides chain-of-evidence assessment to optimize reviewer effort by risk level
Stack: Vercel Functions, Vercel Queues, Vercel Sandbox, Neon Postgres; GitHub webhooks trigger agent runs
Case study: feature request issue #17898 (blocked-domain support) processed end-to-end by factory agents in single workflow, including backports to v5 and v6
Go to the source
Vercel Blogvercel.com
Publisher excerpt: The AI SDK is one of the most popular open-source AI projects in the world. It serves over 20 million npm downloads a week and the repo has over 26,000 stars. Maintaining the codebase means tracking four moving targets at once: After multiple years of growth, the repo was getting 100+ new issues…