multi-agent-patterns

Create a central supervisor agent that decomposes tasks into subtasks, routes them to specialized workers based on capabilities, and aggregates results. Constrain worker output schemas to prevent supervisor context saturation, and cap workers at 3-5 per supervisor.

Supervisor patterns use a central coordinator that delegates to workers and synthesizes results, giving strict control but creating a bottleneck. Swarm patterns let agents hand off control directly to each other, removing the single point of failure but requiring explicit handoff protocols and convergence checks.

LangGraph uses graph-based state machines with explicit nodes and edges for deterministic routing. AutoGen uses conversational GroupChat patterns for event-driven coordination. CrewAI uses role-based hierarchical crews. Choose based on whether you need explicit state control, conversational flexibility, or role-based structure.

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

LLMs have an inherent bias toward agreement, so agents converge on agreeable rather than correct answers, and simple majority voting weights hallucinations equally with sound reasoning. Counter this with confidence-weighted voting, explicit adversarial roles, and required statements of disagreement before convergence.

Avoid multi-agent designs when a single agent's context window can hold the task, when subtasks cannot run in parallel, or when decomposition creates more coordination overhead than the work itself. Over-decomposition past 3-5 agents shows diminishing returns and quadratic communication growth.