multi-agent-patterns

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

Use multi-agent designs when a single context window cannot hold all task-relevant information, when subtasks run in parallel, or when subtasks need different tools and system prompts. Avoid them for simple tasks where coordination overhead exceeds the benefit.

LangGraph uses graph-based state machines with explicit nodes and edges, AutoGen uses conversational GroupChat patterns, and CrewAI uses role-based hierarchical crews. Choose based on whether you need explicit state control, event-driven conversation, or role-based processes.

Multi-agent runs can cost roughly 15x a single-agent baseline due to coordination messages, retries, and consensus rounds. Budget for this multiplier and measure against single-agent baselines before assuming extra agents improve outcomes.

Avoid simple majority voting, which treats hallucinations equally with sound reasoning. Use confidence-weighted voting, structured debate protocols with adversarial critique, and require agents to state disagreements before convergence is allowed.

The supervisor context becomes a bottleneck as worker count grows, supervisor failures cascade to all workers, and paraphrasing sub-agent responses causes telephone game information loss. Use forward_message tools or swarm handoffs to let workers respond directly.