multi-agent-patterns

Deploy a central supervisor agent that decomposes user objectives into subtasks, routes them to specialized workers, and aggregates results. Constrain worker output schemas to prevent supervisor context saturation, and cap workers at 3-5 per supervisor.

Supervisor architectures use a central coordinator that delegates to specialists, giving strict control but creating a bottleneck. Swarm architectures let any agent hand off control to any other peer, offering flexibility and no single point of failure at the cost of harder convergence.

LangGraph uses graph-based state machines with explicit nodes and edges for deterministic workflows. AutoGen uses conversational GroupChat patterns for event-driven coordination. CrewAI uses role-based hierarchical crews for structured team processes.

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.

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 and explicit adversarial roles.

Avoid multi-agent designs when a single agent's context window can hold all task-relevant information, or when subtasks do not benefit from isolation. Over-decomposition creates more coordination overhead than the task itself, and returns diminish past 3-5 agents.