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 the supervisor context from becoming a bottleneck, 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 directly to each other, eliminating translation errors but requiring explicit handoff protocols and convergence constraints.

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 the 15x multiplier including coordination overhead, retries, and consensus rounds.

Yes, LangGraph implements supervisors as graph-based state machines with explicit nodes and edges. The references include code for supervisor nodes that route to researcher and writer workers using StateGraph with typed state.

LLMs have an inherent bias toward agreement, causing sycophantic consensus where agents mimic each other rather than reason independently. Counter this with weighted voting by confidence, explicit adversarial roles, and requiring stated disagreements before 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 into many agents spends more tokens on handoffs than on actual work.