Agent On Rails
Agent On Rails
From specs to running software.
Define the contract. Let agents deliver it — with review, escalation, and evidence before done.
Agent On Rails
Agent On Rails
Define the contract. Let agents deliver it — with review, escalation, and evidence before done.
How it works
Agent On Rails does not start by running an agent. It starts by making the product contract real — then governing execution until evidence proves completion.
Product docs, ADRs, policies, and specs live in the control plane.
Approved specs become a task graph — not one unbounded prompt.
An AI Team implements behind PRs with cheap-first model routing.
Independent review + evidence, then human final review before merge.
AI Team
Define Manager, Implementor, and Reviewer roles — then let Agent On Rails operate the team until the contract is satisfied.
Orchestrates next steps, retries, and escalation within policy.
Writes and fixes code against the bounded task contract.
Judges diffs against acceptance and evidence — never self-approves.
Why it exists
Approved specifications bind planners, agents, and reviewers.
Completion requires PRs, tests, review, and an audit trail.
Start inexpensive. Raise model tier only when the task demands it.
Issues, labels, PRs, and Checks are the operating surface.
Dogfood
Agent On Rails is being built using Agent On Rails. The first external proving ground is Lakuyo — a multi-repo POS delivery system already running on control-plane discipline.
The control plane is the source of truth — product docs, ADRs, policies, specs, and acceptance criteria.
Read the how-to guide