The first AI agent you build will want to do everything itself. One model, one loop, one giant tool list, trying to answer "what's next?" every tick. This works for toy problems. It breaks at about five actions in.
Professional kitchens figured out a better way 150 years ago.
The brigade system — invented by Escoffier in the 1880s — divided a kitchen into specialised stations, each with a clear owner, a clear scope, and a shared system for coordinating. The saucier makes sauces. The garde manger does cold. The expediter runs the pass. Nobody does everything. Everyone knows who does what.
Your AI agent should have the same shape.
Why "one agent with all the tools" fails
"Just give the model all the tools and let it decide" is a seductive pitch. It's simple to build. It's close to how the demos work. It's what most of the "agent" frameworks assume.
It also doesn't scale. A single model with 40 tools is a single model with 40 tools: each prompt carries the description of all of them, each decision has 40 options, and every tool call is a chance for the model to pick badly. Latency compounds, token costs compound, failures compound. The agent works great for the first three steps of a five-step task — and then drifts.
The instinct to "just use a bigger model" or "just add a better system prompt" is how the first line cook became a single-station kitchen. It doesn't work. Escoffier didn't hire bigger cooks.
What brigades and good agents share
Four structural features:
Specialised stations. A brigade doesn't have one generalist doing pasta, fish, and desserts. It has a pasta station, a poissonnier, a pastry chef. Each has a tight tool set, a clear set of prompts, a scope they don't stray from. Agents designed this way — a "research" agent with web and RAG tools, a "writer" agent with composition tools, an "editor" agent with evals — produce cleaner, cheaper, more debuggable outputs. Each station has five tools, not fifty.
The pass is a protocol. In a brigade, the pass is the counter where everything is coordinated. Plates leave the pass; they don't move without the expediter's signal. In an agent system, the pass is a structured handoff format — a small, typed interface between stations. It's boring. It's rigid. That's why it works. Stations don't negotiate; they conform to the interface.
The expediter is the only generalist. A brigade has one role that sees the whole flow: the expediter at the pass. They don't cook. They coordinate. In an agent system, this is your planner or orchestrator — a small, cheap model that calls the specialists in the right order, checks their output, and routes. They don't use the specialist tools; they use the stations themselves as tools.
Signals, not arguments. A busy kitchen communicates in short, pre-agreed signals: "oui, chef," "behind you," "plating three." Not paragraphs. Brigade-shaped agents work the same way: stations emit small structured events, the orchestrator consumes them. Verbose agent-to-agent prose is the equivalent of a station chef narrating to the expediter why they chose rosemary. It's a smell.
Four design moves
Make stations, not megaprompts. Three agents with 5 tools each outperform one agent with 15 tools nine times out of ten. The reason isn't reasoning depth — it's attention and cost. Smaller tool lists, tighter scope, cleaner behaviour.
The pass is a schema. Define the handoff format before you define the stations. Use strict types: Pydantic, Zod, JSON Schema. If a station has to parse freeform text to understand what to do, you've moved the complexity to the wrong place.
Use cheap orchestrators. Your expediter doesn't need to write code or summarise a document — it needs to pick the next station and route. Use the smallest, fastest model you can. Burn good money on the specialists.
Measure per-station. Evals belong at the station level, not the whole-agent level. If the sauce is off, you don't re-taste the entire menu. Test the saucier. Run station-level evals on every deploy.
Division of labour, plus protocol
There's a reason every serious kitchen in the world uses some version of the brigade. It's not because it's old. It's because division of labour, plus a coordination protocol, scales further than any amount of genius.
Agent design is still arguing with that lesson. We don't have to.
Related reading
- MCP servers are USB-C for AI — the protocol that makes tool-calling portable across stations.
- Prompts are recipes, not spells — what the pass schema between agents actually looks like in practice.
- Your AI coding assistant is a midwife, not a genius — the same handing-back-agency principle applied to dev tools.
We build AI-enabled software and help teams put AI to work. If you're designing agent systems, we'd love to hear about it. Get in touch.