Why Agent Brigades

Part 1 of 3 — the concept. For anyone. (Part 2: the anatomy, for engineers. Part 3: running one, for delivery teams.)

The problem this solves

A single AI agent given a big job produces output nobody reviewed by construction. It writes the spec, satisfies its own spec, grades its own work, and hands you something polished-looking whose defects are invisible precisely because the same context produced and judged them. At delivery scale — client assessments, sales collateral, production skills — that's not a quality problem, it's a certification problem: you can't tell the org "this artifact is good" when the only witness is the author.

The fix is old. Professional kitchens solved it in the 1890s: the brigade. Not one heroic cook, but named stations with narrow jobs, a pass where a decider stands between the kitchen and the dining room, and an order ticket that travels with the dish and records what happened to it. Escoffier's insight was that quality at volume comes from role separation plus a contract at every handoff — not from hiring a better cook.

An Agent Brigade is that pattern applied to AI agents.

The shape in one paragraph

Work enters as a ticket hung on a rail (a queue with a paper trail). An expo pulls the ticket, checks it's sufficient, and routes it through stations — each a separate agent with one narrow job, handing off through file artifacts rather than shared context. Independent critics evaluate the result and advise; the expo decides — advance it, refire it, reroute it, or kill it. Approved artifacts land in the cellar (the store of record), and the ticket — now carrying every verdict, every refire, every human approval — files away as the permanent build record.

The three seams that do the work

Everything else is plumbing; these three separations are the pattern:

1. The critic advises; the expo decides. Evaluation and routing are different jobs. Critics are spawned with zero context on the build — they can't be anchored by the author's reasoning. But a verdict is not a route: the expo holds information critics don't (round budgets, ticket history, what refiring costs) and owns the exit.
2. The test author is blind to the implementation. Whoever defines "what success looks like" reads only the spec — never the code. Same reason auditors don't report to the CFO.
3. The front of house never cooks. Requesters and the person managing intake (the steward) live outside the kitchen. The only thing that crosses the boundary is the ticket. That means the kitchen can run anywhere — a laptop, a headless box, a container — without the intake surface caring.

"Isn't this just a swarm?"

Swarms — many peer agents coordinating emergently — are genuinely good at exploration: search, brainstorming, coverage of an unknown space. The brigade is the opposite bet, for the opposite moment: when the output must be certified, you want fixed roles, contracts at boundaries, and an audit trail, because emergence is precisely what you can't put in front of a client. The two compose: several brigades in this house fan out swarm-style inside a station (one worked run pushed ~90 parallel evaluations through sub-agents without touching the deciding agent's context), then collapse back to the ticket. Emergence inside, contracts at the seams.

The principle underneath: the build record IS the product

A brigade artifact is worth more than the same text produced by one prompt, because of what stands behind it: the spec it was built to, the tests defined before the build, the critic verdicts, the human gates it passed, the refires it survived. All of that lives on the ticket. Two consequences worth internalizing:

• "Truth lives on the ticket; there is no second store to reconcile." Status meetings, dashboards, and hand-offs are all views over the same filed record.
• Direct station output — calling one station à la carte, outside the brigade — is treated as un-gated work product. Useful, but never filed as brigade-certified, because "the value of a brigade artifact is the build record behind it; à la carte output has none, and pretending otherwise is how slop gets certified."

What exists today

This is not a proposal. A working house runs now: a skill factory (a brigade whose product is new agent skills — and, recursively, new brigades), domain brigades for company research, assessments, and sales collateral, a website brigade whose artifact is code commits, and a registrar handling intake and records for all of them. The pattern has served real engagements end-to-end — research → assessment → committed deliverables → rendered marketplace — with the governance visibly working: independent reviewers vetoed real drafts mid-run, and the refires produced honestly better output.

Part 2 names every piece precisely. Part 3 shows a full day of running one.