multi-agent-patterns

Deploy a central supervisor agent that decomposes tasks into subtasks, routes them to specialized workers based on capabilities, and aggregates results. Constrain worker output schemas and use checkpointing to prevent the supervisor context from becoming a bottleneck.

Supervisor patterns use a central coordinator that delegates to workers, giving strict control but creating a single point of failure. Swarm patterns let agents hand off control directly to each other, offering flexibility but requiring explicit convergence constraints.

Production data shows multi-agent systems cost approximately 15x the tokens of a single-agent chat, versus roughly 4x for a single agent with tools. Budget for coordination overhead, retries, and consensus rounds, not just per-agent costs.

Avoid simple majority voting, which treats hallucinations equally with sound reasoning. Use weighted voting based on confidence and expertise, assign explicit adversarial roles, and require agents to state disagreements before allowing convergence.

Avoid multi-agent designs when a single agent's context window can hold all task-relevant information, when subtasks cannot run in parallel, or when coordination overhead exceeds the work itself. Over-decomposition creates more handoff cost than value.

Use retry logic with exponential backoff and circuit breakers that temporarily disable failing agents after repeated errors. Validate outputs before passing them downstream, and reroute tasks to backup agents when circuit breakers activate.