Multi Agent Patterns

When you hit context limits on complex tasks, this walks you through splitting work across multiple Claude instances. It covers three patterns: supervisor (central coordinator delegates to specialists), peer-to-peer (agents hand off directly via function returns), and hierarchical (layered abstraction). The big insight here is that multi-agent systems aren't about simulating teams, they're about context isolation. Fair warning though, token costs jump to roughly 15x compared to single-agent approaches, so you're trading money for capability. The skill includes a practical fix for the "telephone game" problem where supervisors mangle sub-agent responses, using a forward_message pattern that lets specialists respond directly to users.