Multi Agent Patterns

This breaks down when and how to split work across multiple Claude instances instead of cramming everything into one context window. The core insight is that multi-agent patterns are mainly about context isolation, not role-playing organizational charts. You get three main architectures: supervisor/orchestrator for centralized control, peer-to-peer swarm for flexible handoffs, and hierarchical for layered work. The token economics are brutal though. Multi-agent systems can burn 15x more tokens than single-agent approaches, so this makes sense when you're actually hitting context limits or need genuine parallelization, not just because it sounds sophisticated. The guide includes specific fixes for common problems like the "telephone game" where supervisors mangle sub-agent responses.