Multi-Agent Coordination Patterns

A coordination pattern describes the relationships between participants rather than the capabilities of an individual model. In a supervisor pattern, one agent delegates and combines results; in a handoff, responsibility passes to another participant; in a parallel pattern, independent branches run before their results are merged. Voting or critic roles introduce different decision rules. These arrangements determine who sees which context, who can call tools and how disagreements are handled. The pattern is distinct from orchestration, which implements scheduling and lifecycle behavior. Labels such as swarm do not specify reliable coordination unless messages, authority and termination are defined.

The practitioner maps task dependencies and chooses roles only where separation provides a benefit. They define input and output contracts, context-sharing rules and a final decision owner. Parallel branches need a merge policy, while handoffs need a clear transfer of responsibility. A useful design includes a topology and test cases for conflicting or incomplete results. Comparison with a single-agent baseline checks whether the additional participants improve coverage or specialization enough to justify their communication overhead.

Extra agents can duplicate work, propagate one another's mistakes or create unclear ownership. Voting among closely related models does not provide independent evidence. A hierarchical manager may become a bottleneck, while unrestricted peer communication can be difficult to debug. Quality depends on task-appropriate separation, explicit contracts and an enforceable stopping rule. The number of roles is not a measure of sophistication; coordination is useful when it solves a concrete dependency or information-boundary problem.