Multiple Cooks, One Kitchen

Opening Scene

A genuinely complex kitchen doesn’t rely on one cook doing everything. It has a grill station, a sauté station, a pastry station, each staffed by someone with focused expertise, all coordinated by the expediter into a single coherent service. Some agentic tasks are similarly better handled by several coordinating agents with distinct, focused roles than by one single, generalist agent trying to do the entire task alone.

In Plain English

Multi-agent systems split a complex task across several agents, each with a distinct, focused role, coordinated by some orchestrating structure so their individual outputs combine into a single coherent result. This isn’t automatically better than a single generalist agent — it adds real coordination complexity — but for genuinely complex tasks with distinct, separable subtasks, focused specialist agents can each perform better within their narrower scope than one generalist agent attempting the entire task alone.

The Old Way

Before multi-agent systems became a well-established, deliberately chosen pattern, early agentic approaches often defaulted to a single agent regardless of task complexity:

Early agentic systems often used a single, generalist agent for genuinely complex tasks, even when the task had clearly separable subtasks that focused specialists could plausibly have handled better individually.
There wasn’t yet well-established tooling or orchestration patterns for coordinating multiple agents’ outputs into one coherent result.
The added coordination complexity of multi-agent systems wasn’t yet well understood, making it genuinely hard to know when the approach was actually worth its overhead.

Multi-agent systems emerged specifically as agentic tasks grew complex enough that a single generalist agent’s limitations became a genuine, practical bottleneck.

What’s Changing (and Why AI Is the Reason)

Frameworks and orchestration patterns for coordinating multiple specialist agents have matured significantly, making multi-agent systems a genuinely practical option, not just a research curiosity.
Well-designed multi-agent systems increasingly assign focused, narrow roles to individual agents — connecting directly to the tool selection and scoping principles covered in Article 6 — rather than simply running several generalist agents in parallel.
This connects directly to the parallel execution pattern covered in Article 11, since multi-agent systems often rely on the same underlying ability to run independent subtasks concurrently.

The Metaphor, Fully Extended

The Kitchen	Multi-Agent Concept
Grill, sauté, and pastry stations, each with focused expertise	Specialist agents, each with a distinct, focused role
The expediter coordinating each station into one coherent service	An orchestrating structure combining agents’ outputs into one result
More stations meaning more coordination overhead to manage	More agents meaning more coordination complexity to manage
A kitchen built with enough separable stations to genuinely justify the structure	A task with clearly separable subtasks that genuinely justify a multi-agent approach

For Beginners: What to Actually Do

Practice identifying whether a given complex task has clearly separable subtasks that focused specialist agents could plausibly handle better than one generalist.
Learn to trace how a simple two-agent system’s outputs get combined by whatever orchestrating structure coordinates them.
Get comfortable recognizing that multi-agent systems add real coordination overhead, and aren’t automatically the better choice for every complex task.

For Practitioners and Leaders: The Deeper Layer

Evaluate multi-agent architectures specifically for tasks with clearly separable subtasks, rather than defaulting to them for all complex problems.
Design individual agents’ roles to be genuinely narrow and focused, connecting directly to the tool scoping principles covered in Article 6.
Weigh the real coordination overhead multi-agent systems introduce against the potential performance gains from specialization.

Quick Recap

Multi-agent systems split a complex task across several agents with distinct, focused roles, coordinated into one result.
This adds real coordination complexity, and isn’t automatically better than a single generalist agent.
It’s most justified for tasks with clearly separable subtasks that focused specialists can handle better individually.
Orchestration patterns and frameworks for coordinating multiple agents have matured significantly.

Where This Fits in the Series

Article 16 covered coordinating multiple agents within one system. Article 17 covers what happens when an agentic system goes wrong, and the guardrails that keep a failure contained.