Building a software factory for AI SDK
By Steven Van ·
The AI SDK is one of the most popular open-source AI projects in the world.
Vercel built an internal system called ai-sdk-factory to work through the backlog on its AI SDK open-source project. By late June the repo, which sees over 20 million npm downloads a week, had accumulated more than 1,000 open issues and almost 800 pull requests, after months of 100+ new issues a month and a spike following Anthropic's Opus 4.6 release. Four weeks after deploying the factory, it authors between 25 and 35% of the PRs the team merges and closes 70 to 80% of issues.
The factory runs a set of dedicated agents, each scoped to one task in the pipeline:
- Bug reproduction
- Bug fixes
- PR reviews
- Backports
- Documentation updates
- Feature analysis
- Feature implementation
Every agent runs inside an isolated Vercel Sandbox with only the secrets its task needs, behind a network shielding layer meant to block attempts to exfiltrate secrets or prompts from untrusted issue and PR content. No change ships without a human on the AI SDK team reviewing and merging it. The system itself runs on Vercel Functions, Vercel Queues, Vercel Blob, Vercel Sandbox, and Neon Postgres, with a monitoring UI that shows reviewers every run in the pipeline.
Read the full write-up in the original announcement.
