Multi Agent Supervisor Orchestrate

Designs supervisor/worker graph topologies where a coordinating Claude instance dispatches work to specialized subagents and integrates their results. Covers decomposition criteria, the supervisor's dispatch-collect-decide loop, context digesting to prevent state bloat, and failure isolation between workers. TRIGGER when: designing a multi-agent system with a coordinator and specialized subagents, deciding whether to split one agent's tool set across multiple agents, debugging a supervisor whose context window fills up or whose decisions degrade over long runs, or handling a subagent that returns bad/hallucinated output. DO NOT TRIGGER when: building a single agent with a…