When (and When Not) to Go Multi-Agent
The real costs and benefits of multi-agent systems, and a checklist for deciding if your problem needs more than one agent.
"Multi-agent" is one of the most hyped ideas in AI right now. Diagrams of a CEO agent managing a team of specialist agents look impressive, and demos of ten agents working at once are genuinely fun to watch. They also burn money, fail in confusing ways, and often produce worse results than one well-equipped agent would have.
This module teaches you to build multi-agent systems that actually earn their complexity. This first lesson is the most important one: deciding whether you need more than one agent at all. You'll learn what multiple agents really buy you, what they really cost, which problems fit and which don't, and you'll leave with a checklist you can apply to any project before writing a line of orchestration code.
What "multi-agent" actually means
Strip away the org-chart metaphors and a multi-agent system is simple: more than one agent loop, each with its own context window, cooperating on a task. You built one agent loop in the previous module. A multi-agent system runs several of those loops and passes information between them.
Four patterns cover nearly everything you'll see, and each gets its own lesson in this module:
| Pattern | Shape | Example |
|---|---|---|
| Orchestrator–worker | A lead agent splits a task, workers do the pieces, the lead combines results | Research across several sources |
| Subagents | A main agent delegates a self-contained chunk and gets back a short report | "Read these 40 files and list the TODOs" |
| Parallel fan-out | The same kind of work runs on many items at once | Summarize 200 support tickets |
| Pipeline with handoffs | Agents work in sequence, each passing an artifact to the next | Research, then write, then review |
These patterns combine. An orchestrator usually fans out to its workers in parallel, and a pipeline stage can use subagents. But each one rests on the same basic move: give a piece of work to a fresh context and bring back a result.