2.3. Multi-Agent Collaboration

💡 First Principle: Agents scale the way teams do — by delegation, not by growth. Past a point, adding more tools and topics to one agent degrades its routing (too many similar descriptions competing for the orchestrator's attention); the cure is specialists: a main agent that understands intent and hands work to agents that each do one thing well.

Copilot Studio expresses this with a parent that delegates to child agents (lightweight subagents inside the same agent, sharing its configuration and authentication) and connected agents (independent agents — other Copilot Studio agents, Foundry agents, Fabric data agents, or external A2A agents — each with its own lifecycle). The delegation mechanics are consistent: the main agent's orchestrator picks the right collaborator for the query, the collaborator runs its own orchestration with its own tools, and results flow back.

⚠️ Common Misconception: Multi-agent collaboration is just an agent calling a flow. A flow is a deterministic procedure; a delegated agent reasons — it has instructions, tools, and knowledge of its own and can conduct its own multi-turn work before reporting back.