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.

Use supervisor patterns when tasks have clear decomposition and human oversight matters. Use swarm/peer-to-peer patterns when requirements emerge dynamically and sub-agents can respond directly to users, which eliminates telephone-game translation errors from supervisor paraphrasing.

Production data shows multi-agent systems cost approximately 15x the tokens of single-agent chat, while single agents with tools run about 4x baseline. Budget for 15x and treat lower costs as a bonus, since coordination overhead and retries add up.

Yes, the references directory includes framework-specific implementations. LangGraph uses graph-based state machines with explicit nodes and edges, AutoGen uses conversational GroupChat patterns, and CrewAI uses role-based hierarchical crew structures.

LLMs have an inherent bias toward agreement, causing sycophantic consensus where agents agree to be agreeable rather than correct. Counter this with weighted voting by confidence, explicit adversarial roles, and requiring agents to state disagreements before convergence.

Adding agents past 3-5 shows diminishing returns because communication channels grow quadratically and coordination overhead exceeds parallelization gains. Start with the minimum viable agents and add only when clear context isolation benefits exist.