Build a complex agent from the ones already in the hub. Give it a set of member agents (and optionally your own MCP + prompt); on each request a planner decides whether to route to one agent, run a sequential pipeline, or fan out to agents in parallel and merge their answers — and you watch every decision.