Choosing a pattern

Each multi-agent pattern gives up something to gain something. Pick by what the task needs, not by which pattern sounds most capable.

Pattern	Shape	Gains	Costs	Best for
Router	Manager picks a specialist	Maintainable, testable, small calls	Depends on accurate classification	Requests that fall into clear categories
Handoff	Fixed sequence	Deterministic, reliable, simple	Rigid, no backtracking	A known pipeline, such as research, draft, review
Group chat	Agents discuss in one thread	High quality, flexible, adaptable	Expensive, slow, non-deterministic	Decisions that need debate or review
Swarm	Queen spawns parallel workers	Massive speed and scale	Expensive, parallel jobs only, complex	Many independent items, such as 10,000 articles

Reading the table

The two ends of the table pull opposite ways. Handoff is the predictable one and group chat is the flexible one, and flexibility is paid for in cost, latency and repeatability. The routerRouter patternA manager that only decides which specialist handles the request.Opens in a new tab sits between them, and the swarmSwarm and map-reduceA queen agent spawns a worker per task, then collects the results.Opens in a new tab is the odd one out because it changes how fast the work finishes, not how it is decided. See handoffHandoff patternAgents in a fixed sequence, each passing its output to the next.Opens in a new tab and group chatGroup chat patternAgents that talk to each other in one thread, with a manager keeping order.Opens in a new tab for the details.

Before any of them

A single agent with a handful of tools is often enough, and multi-agent systems cost far more tokens (see the monolith agent problemThe monolith agent problemOne super agent that does everything breaks the way a software monolith does.Opens in a new tab). Reach for a pattern when one agent is overloaded with tools, prompts or context.

The patterns also combine. A router can sit in front of a sequence, a group chat can be one stage of a pipeline, and a swarm can be one node in a larger graph. LangChain’s docs describe this as a custom workflow, and Microsoft’s guide has a section on combining patterns. Whatever the mix, add a human checkpointHuman in the loopA pause in the agent's decision, where a person approves before a high-risk action runs.Opens in a new tab where an action is hard to undo, and traceTraceabilityLog the whole reasoning trace, so any answer can be explained afterwards.Opens in a new tab every agent so a bad run can be explained.