About This Skill

The hard part of multi-agent work isn't spawning agents — it's deciding what deserves parallelism, what each agent needs to not duplicate work, and how failures cascade when one agent in a pipeline returns bad output. This skill designs the full orchestration: which tasks benefit from parallelism vs. which are serial dependencies, what context each subagent needs and nothing more, how results are validated before being handed downstream, and where human checkpoints belong. It maps your workflow into an agent graph with clear boundaries, failure modes, and fallback strategies — before you write a single tool call. Use it when you're building multi-step Claude pipelines, Claude Code subagent flows, or any system where one model's output feeds another. Give it your goal, your available tools, and your constraints; it returns an architecture you can implement without discovering the coordination problems at runtime.