One agent becomes a team. An orchestrator (coordinator) decomposes a goal, delegates to specialist subagents, then aggregates and evaluates their results. Specialists get narrow tools and isolated context — a single agent with 30 tools and 10 responsibilities produces worse output than a team of specialists.

The core pattern is hub-and-spoke: coordinator at the center, subagents as spokes. Key decisions: parallel vs sequential execution, structured handoffs that preserve source attribution, and knowing when a single agent is actually the right call (reliability before autonomy).

📚 Resources

• 🎬 Explainer — Why one super agent fails: Agentic AI orchestration explained — orchestrator + specialist agents in a real pharma platform, why reliability comes before autonomy, and where humans stay in the loop.
• 🛠️ Build-along — How to Build a Multi-Agent AI System (Step-by-Step) — supervisor + worker sub-agents with the Architect-Builder-Critic framework, memory isolation, and the 4 golden rules of multi-agent systems.
• 🏭 In the wild — coding-agent swarms, research pipelines with parallel subagents, and this exact workflow: one coordinator delegating research while you stay in the chat.