Vercel Labs released Foreman, which runs tasks through four agent stations

Vercel Labs published vercel-labs/eve-software-factory-template on 12 August 2026. It released a template called Foreman, described in the README as an eve software factory that puts AI agents on every stage of the development loop and keeps people on the judgment calls. The repository has 1,093 stars and 60 forks in its first 22 days on GitHub.

What Foreman actually does

Foreman takes a task from GitHub or Linear and moves it through four stations before opening a draft pull request. The README names each station and what it is responsible for.

A Classifier triages the task, deciding type, priority, complexity, and whether it is actionable at all. If a task is not actionable, Foreman asks the requester rather than building the wrong thing. An Analyst turns the request into a plan with acceptance criteria, working from a live checkout of the repository. An Implementer executes the plan in its own sandbox, runs the repository's own checks, and pushes a branch. A Reviewer independently judges the work against the real diff, with evidence attached to each verdict.

Each station is its own agent with its own instructions, sandbox, and tools. The README calls out that the Reviewer only sees the pushed branch and never the Implementer's reasoning, which is the isolation the template relies on for the Reviewer to be an independent check rather than an approval given without real checking. Between runs, Foreman keeps what the README calls a factory brain, a set of notes about the repository that every new run starts from.

How work reaches it

Foreman accepts work from five entry points, all documented in the README. Labelling an issue with factory hands it to the pipeline, which posts progress as each station completes and ends with a draft pull request linked to the issue. An @mention on an issue or pull request from a repository owner, member, or collaborator starts an interactive session. Delegating in Linear runs the same pipeline through Linear Agent Sessions. A local terminal UI accepts a task and waits for approval before it touches GitHub. Failing CI on a factory pull request triggers Foreman to push a fix to its own branches, never to yours.

Deployment

Foreman deploys from the Vercel template flow, which the README says sets up the GitHub and Linear connectors, a Vercel Blob store, and prompts for FACTORY_REPO and FACTORY_LABEL. The FACTORY_REPO value has to point at a repository the GitHub App can reach, and the deploy warms up the station sandboxes by cloning it in advance, so an unreachable repository fails the deploy with a clear error message.

What moves between stations is finished work, never reasoning

Foreman is a specific answer to a real question about how far a single coding agent can go before quality drops. Splitting the work across four narrow agents that hand each other finished work, and never each other's reasoning, is the pattern most teams have to build themselves the first time they take an agent beyond a small task. Vercel Labs releasing it as a deployable template shortens the distance between "we should try this" and "we are running it against our real backlog".

The pattern also has a real limit that the README does not hide. Every station is still a language model making judgment calls, and the Reviewer's independence is only as strong as the split between the branch and the plan the Implementer wrote. A team adopting the template inherits both the discipline and the failure mode, and the useful test on day one is whether the Reviewer catches a plausible-looking wrong diff.

Source

• vercel-labs/eve-software-factory-template on GitHub