Overview

factory is a runtime for self-improving agentic loops. Code is the first product line.

You already have a coding agent that can write a patch (Claude Code, Gemini / Antigravity, OpenAI Codex, Cursor, Pi). What you probably do not have is the process around it:

• Something that decides which work is actually ready,
• Hands one agent one ticket in one isolated worktree,
• Re-runs the verification command independently, and
• Holds the merge until CI and a reviewer agree.

That is factory.

Core philosophy

The tracker is the control plane, git holds the product truth, and CI is the reward signal. Nothing merges because an agent said it was done; it merges because the tests passed and a reviewer approved.

Key Capabilities

Capability	How Factory Delivers It
One ticket, one worktree	Every dispatched ticket runs in a dedicated git worktree with isolated ports, databases, and temporary state.
Concurrency without collision	Dispatched tasks require disjoint Owned Paths globs. The planner prevents two agents from editing the same files simultaneously.
Independent Verification	The factory worker re-runs the declared verification command. An agent’s report is commentary; the test exit code is evidence.
No product truth locked in Factory	Code lives in git and tasks live in Linear or GitHub Issues. Runtime state and receipts can be rebuilt from those durable records.
Self-improving loops	Defects, friction, and test failures are turned into issues, prioritized, and dispatched to maintain the factory itself.

Next Steps

• Connect Your Repo — Point factory at a repository you already have. The fast path is a prompt you hand to the coding agent you already use.
• Your First Dispatch — What happens when factory picks up a ticket, and how to read the result.
• Configuration Reference — Repositories, policies, and schedules.
• Factory Doctor — Verify your local environment and toolchains, or check a fresh clone offline.