Делегирование задач дочерним агентам

orchestration

agent-orchestration is a Claude Code skill that defines a delegation pattern for complex implementation tasks, keeping the main context clean. Instead of reading files and making edits in the main chat — which consumes 2000+ tokens — Claude spawns an implementation agent with a clear task and receives only a concise summary (~200 tokens) in return. The skill specifies when to use agents: multi-file implementations, following a plan phase, building new features with tests — and when not to bother: a single-line fix or a quick config change where the overhead isn't worth it. A core principle is that agents don't need files pre-read for them; they figure out the context themselves. The skill triggers automatically on words like "implement", "build", "follow the plan", or "do phase X".