Multi Agent Patterns

This breaks down when to split work across multiple Claude invocations instead of cramming everything into one context. The key insight here is that sub-agents exist to isolate context, not to play organizational theater. You get three patterns: supervisor/orchestrator for centralized control, peer-to-peer for flexible handoffs, and hierarchical for layered abstraction. Watch out for the telephone game problem where supervisors mangle sub-agent responses when paraphrasing. The file system becomes your coordination mechanism. Use this when you're hitting context limits, when subtasks can run in parallel, or when specialized prompts would improve quality over a generalist approach.