eve agents that code from live docs
In this example, Vercel's eve Software Factory reaches the web through the @nimble-way/eve extension at every station, and turns a GitHub issue into a reviewed draft pull request. For a dependency upgrade, Nimble Extract reads the official guide for the exact version pair, with Nimble Search finding it when needed, so the plan follows today's docs rather than the model's training data. The factory then edits the code in a sandbox, runs the repository's checks, has a second model review the diff, and opens a draft PR that links its sources. The demo upgrades a small Next.js app from AI SDK 6 to 7.
Inputs
- A repository Any GitHub repo the factory's GitHub App can reach. The cookbook ships a demo app pinned to AI SDK 6.
- An issue Labeled factory, for example "Upgrade to AI SDK 7: follow the official migration guide, keep behavior the same, checks must pass."
Outputs
- A draft pull request with the migration, passing the repository's checks
- A Sources section linking the guide pages the plan relied on
- Progress comments on the issue as each station finishes
- Each run in Vercel Observability, including every Nimble call
Draft PR #7 on a scratch copy of the demo app records one full run: the AI SDK 6 to 7 upgrade with acceptance criteria, passing typecheck, tests, and build, an approved review, and a Sources section citing the official migration guide and four related pages. That run made 33 web calls across its stations, all through Nimble.
How it works
Five eve stations, from labeled issue to draft PR.
- Classify the issue A fast model triages it: actionable or not, complexity, and whether another open issue or PR already covers it.
- Research with Nimble The researcher reads the official migration guide for the exact version pair with Nimble Extract, finding it with Nimble Search when needed. Every finding cites the guide, and later stations check details through the same Nimble tools.
- Plan from the guide The analyst writes a plan with acceptance criteria from the findings and a live checkout. Behavior changes a codemod misses, like AI SDK 7's usage now totaling every step, become criteria.
- Implement in a sandbox The implementer edits the code in its own Vercel Sandbox, runs the repository's typecheck, tests and build, and pushes a factory branch.
- Review and open a draft PR A reviewer on another vendor checks the diff against the criteria. The draft PR lists every source URL, and merging stays with a person.
Stack
- @nimble-way/eve The eve extension, behind every agent's web tools. Its Web Search Agent tools are disabled so an unattended run never waits on an approval.
- search Replaces the built-in web_search in every station, to find official, version-specific docs.
- extract Replaces the built-in web_fetch in every station, to read those pages in full as clean markdown.
- eve 1.0 Durable orchestrator, subagent stations, GitHub channel.
- Vercel Sandbox Isolated checkout where code is edited and checks run.
- openai/gpt-5.6-terra-fast Classifier, researcher, analyst, reviewer, through Vercel AI Gateway.
- anthropic/claude-fable-5 Implementer, on a different vendor than the reviewer.
- Vercel Connect and Blob GitHub App and the factory's memory.
