multi-agent-patterns

Create a central supervisor agent that decomposes tasks into subtasks, registers workers by capability, and routes work through a message channel. The SupervisorAgent class in the coordination script handles worker selection, task assignment, and result aggregation.

Supervisor patterns centralize control in one coordinator that delegates to specialists, while swarm patterns let any agent hand off control to any other through explicit transfer functions. Choose supervisors for clear task decomposition and swarms for flexible, emergent exploration.

Implement a forward_message tool that lets sub-agents respond directly to users instead of routing through supervisor synthesis. Alternatively, use filesystem-based shared state so agents read original data rather than paraphrased summaries.

Yes, the references document includes framework-specific implementations: LangGraph state graphs for supervisors and swarms, AutoGen GroupChat for conversational coordination, and CrewAI-style hierarchical manager-worker delegation.

Production data shows multi-agent systems run at roughly 15x the token cost of single-agent chat due to coordination overhead, retries, and consensus rounds. Budget accordingly and validate that parallelization actually saves wall-clock time.

Avoid multi-agent designs when a single context window can hold the task, when subtasks are not truly parallelizable, or when coordination overhead exceeds the work itself. Adding agents past 3-5 shows diminishing returns and quadratic communication growth.