Pipeline

Issues in. Merged pull requests out. Every step on record.

Nine specialized agents plus a Code Map, coordinated by a deterministic state machine. Each agent does one job and records the evidence for it. If anything stops, the next run picks up exactly where the issue sits — not in someone’s head.

Two weeks to first merged PR.

The lifecycle agents

One agent. One job. Its own evidence.

No agent is a generalist. Each owns one step and the evidence that justifies it; Sentinel watches all of them.

01

Mayor

Decides what to work on, and in what order.

Reads the backlog and queues issues by priority, dependency and capacity. Nothing starts out of order, and blocked work waits for what it depends on.

02

Surveyor

Reads the issue, the repo, and the conventions.

Resolves the issue to a concrete problem: affected files, existing patterns, open unknowns. If the issue is ambiguous, it asks a person instead of guessing.

03

Architect

Lays out the change before code is written.

Produces a structured plan — files touched, functions modified, tests added. Work too large for one pull request is split here, before any code exists.

04

Builder

Writes the code.

Implements the plan, runs the tests and opens the pull request. The PR carries its own evidence: test output, diff summary, files touched.

05

Inspector

Catches what would have been merged anyway.

Reviews the diff against your conventions, the plan and the evidence. If the evidence doesn’t hold, the work goes back to the Builder.

06

Marshal

Respects the human-defined gate.

Merges under your branch protection: required reviews, status checks, code-owner approvals. It never elevates its own permissions, and with auto-merge off it hands the call to a person.

07

Sentinel

Watches the pipeline itself.

Tracks throughput, cost and failure patterns across every lane. It surfaces what the other agents can’t see because they only know their own step.

Beyond the lifecycle

Some run on a trigger or a schedule. Not a step.

08

Chronicler

Writes the post-mortem.

After each issue closes, it records what happened and nominates lessons for Repo Intelligence.

09

Mechanic

Repairs failed CI checks.

Makes a bounded, targeted fix attempt when CI fails hard, then hands the issue to a person if it doesn’t hold.

10

Strategist

Plans self-improvement work against an operator charter.

Turns your goals into a weekly plan, and work inside the limits you set ships automatically on Colony’s own repositories — see the Strategist.

11

Code Map

Indexes the repo’s structure. Runs no model.

A deterministic index of files, functions and imports, so the Surveyor can ask what a change touches without reading the whole repo; see Repo Intelligence.

Why specialization works

Generalists improvise. Specialists verify.

Each agent is prompted, tooled and evaluated for exactly one step, and each handoff carries evidence the next step checks. That is why Colony’s failure modes are recoverable instead of mysterious — when something goes wrong, it goes wrong in a named state with a named owner. Read the public failure catalogue.

Big work, handled

An issue too big for one PR becomes a plan, not a bigger PR.

Epics decompose into sub-issues with dependency edges, can build against their own feature branch, and get reviewed as a whole before merge — with an optional second, adversarial review pass. Gaps become new sub-issues; a human signs off at the end.

Operator control

Anyone can step in, from a GitHub comment.

Pause, resume, retry, re-route or reset any issue with a comment; the pipeline acknowledges each command and refuses any transition the workflow does not allow. The dashboard carries the same pause, resume, retry, change-state and cancel controls.

Full command list in the docs

Watch it work before you buy it.

The receipts — live ledger, recent merges, full agent transcripts — are public.