Multi-agent orchestration is a fancy term for one question: how do the agents work together? Three philosophies are fighting it out. - Graphs (LangGraph): nodes and edges, explicit state. Maximum visibility, maximum control. Heaviest when the workflow does not matter. - Roles (CrewAI): agents as team members with jobs, a manager coordinates. Fast to design, fast to explain to a stakeholder. - Conversations (AutoGen): agents talk, work emerges from dialogue. Most flexible, least predictable. - Prototype the same small task in two styles for an afternoon. That teaches more than a week of reading. - Plan observability early: run traces and structured logs before your first overnight run. A monitor app is a luxury; traces are the minimum. Want the agent without the homework? PrivateLLM deploy sets up your private LLM on AWS for $50 plus usage.