Multi-Agent Orchestration
When one model loop becomes a tiny office and why that can help or make everything slower.
An agent is usually an LLM wrapped in a loop that can inspect state, choose actions, call tools, and continue until it decides it is done. Multi-agent orchestration means you run more than one of those loops with different roles.
Common patterns
Planner and worker: one agent breaks work down, another executes.
Researcher and writer: one gathers facts, another turns them into output.
Reviewer: one agent checks another agent's work.
Router: one agent decides which specialist should handle a request.
When it helps
Multi-agent setups help when work naturally splits into roles, when parallel work saves time, or when a review pass catches expensive mistakes.
When it hurts
They hurt when the task is simple, when every agent gets too much context, or when nobody owns the final decision. More agents can mean more latency, more cost, more failure modes, and more confident nonsense.
The grown-up version is not "add agents." It is "define responsibilities, tools, stopping conditions, and evals."