Meet Foreman, a software factory that puts AI agents on every stage of the development loop and keeps people on the judgment calls. Built on eve, Vercel's framework for durable AI agents.
Label a GitHub issue factory, mention the bot on a thread, or delegate an issue in Linear, and the orchestrator moves the work through four stations: a classifier triages it, an analyst plans it against a live checkout of your repository, an implementer writes the code and runs your repo's own checks in an isolated sandbox, and a reviewer on a different model vendor judges the pushed branch against the plan's acceptance criteria, sending it back for up to two revision cycles. The finished product is a reviewed draft pull request on your repository. You review, mark ready, and merge.
How it works
- Four specialist stations. Each is a declared eve subagent with its own instructions, sandbox, and tool surface. The reviewer never sees the implementer's reasoning, only its pushed branch, which keeps the review genuinely independent.
- People keep the judgment calls. Unattended runs top out at a draft pull request: marking one ready waits for a human, merging isn't in the tool surface at all, and direct pushes to the default branch are refused in code. Approval prompts render as comments on GitHub, natively in Linear, and as cards in the dev TUI.
- A factory brain. Shared, durable memory of your repository in Vercel Blob: build quirks, verification gotchas, and recurring review findings recorded on one run and read on the next.
- Self-healing CI. When checks fail on a pull request the factory itself pushed, it diagnoses the failure and pushes a fix, pausing for a person after two attempts. Pull requests people push are never touched.
- No static secrets. GitHub and Linear credentials are brokered by Vercel Connect, git credentials are injected at the sandbox firewall and never enter a sandbox, and Blob and the models authenticate with the project's OIDC token.
- An eval suite included. Routing, safety, and prompt-injection behavior are guarded by eve's eval runner, plus an opt-in eval that runs the whole line for real.
