---
title: Overview
description: Foreman is an eve software factory that puts agents on every stage of the development loop and keeps people on the judgment calls.
type: overview
summary: What Foreman is, how a work item becomes a reviewed draft PR, and where to go next.
related:
  - /docs/getting-started
  - /docs/pipeline
  - /docs/trust-model
---

# Overview



Foreman is an eve software factory: it puts AI agents on every stage of the development loop and keeps people on the judgment calls. Read this page to understand what it does, then [deploy it against your repository](/docs/getting-started).

The orchestrator takes work items from GitHub and Linear, moves each one through four stations, and delivers a reviewed draft pull request on your repository (`FACTORY_REPO`). You review, mark ready, and merge.

## The pipeline

Work arrives when a maintainer labels an issue `factory`, someone mentions `@Foreman`, a user delegates through a Linear Agent Session, CI fails on a factory PR, or you hand the local dev TUI a work item. The [intake page](/docs/intake) covers each surface.

However it arrives, every item moves through the same four stations. Each is a declared eve subagent under `agent/subagents/` with its own instructions, sandbox, and tools; the orchestrator (`agent/instructions.ts`) never writes code itself and never skips a station.

1. **Classifier** triages the item on a fast model: type, priority, complexity, affected area, and whether it's actionable. When the item needs clarification, the pipeline stops and asks instead of building the wrong thing.
2. **Analyst** turns the item into a plan grounded in a read-only checkout of your repository: problem statement, approach, risks, and the acceptance criteria the Reviewer later judges verbatim.
3. **Implementer** codes the plan in its own checkout on the strongest coding model, runs your repo's lint, typecheck, and test commands, and pushes a `factory/*` feature branch.
4. **Reviewer** fetches that branch into a separate checkout, reads the real diff, and judges each acceptance criterion with evidence. It approves, requests changes (up to 2 revision cycles back to the Implementer), or rejects.

When the Reviewer approves, the orchestrator opens a draft PR and reports back on the surface the work came from. Alongside the stations, a shared **factory brain** in Vercel Blob carries durable knowledge of your repository into every run and records learnings back once the work lands. See [memory](/docs/memory).

The review is independent by construction. Stations inherit nothing from the orchestrator, so the Reviewer sees only the pushed branch, never the Implementer's reasoning or working state. It also runs on a different model vendor than the Implementer, so it doesn't share the Implementer's idiom or blind spots.

## The trust ceiling

Unattended runs (triggered without a trusted human watching, like a labeled issue) top out at a draft PR. Marking a PR ready waits for human approval, and merging is not in the tool surface at all, so nothing ships without a person. The [trust model](/docs/trust-model) explains how trust is stamped at dispatch and enforced by per-caller approval policies.

## Features

* Six intake surfaces, from unattended label intake to attended mentions and Linear sessions: [intake](/docs/intake)
* Four stations with a clarification gate and a bounded review loop: [pipeline](/docs/pipeline)
* Trust stamped at dispatch, with approval policies per caller class and draft PRs as the unattended ceiling: [trust model](/docs/trust-model)
* Shared factory memory plus per-user preferences in Vercel Blob: [memory](/docs/memory)
* Evals guarding routing and safety behavior: [evals](/docs/evals)
* Everything customizable through ordinary eve files: [customization](/docs/customization)

## Next steps

* [Getting started](/docs/getting-started): deploy the factory against your repository
* [Pipeline](/docs/pipeline): the four stations in detail
* [Trust model](/docs/trust-model): why unattended runs can open draft PRs but never ship


---

For a semantic overview of all documentation, see [/sitemap.md](/sitemap.md)

For an index of all available documentation, see [/llms.txt](/llms.txt)

For agent-facing discovery, including API and MCP surfaces, see [/agents.md](/agents.md)