In one sentence
Coordinating multiple specialized agents when distinct roles and handoffs improve a complex workflow.
Multi-agent orchestration coordinates several AI agents, each assigned a narrower role, through a defined workflow. One agent might classify input, another retrieve approved records, and another assemble a draft for review.
The design can make responsibilities and evaluation clearer when the roles genuinely differ. It does not guarantee better accuracy. More agents introduce more prompts, handoffs, model calls, latency, cost, and failure paths. A single well-scoped agent or deterministic workflow is often the better starting point.
Production systems need explicit tool permissions, state handling, timeouts, retries, traceability, evaluation, and human escalation. The automation boundary and escalation rate should be established with representative test cases rather than an assumed percentage.
My open-source Maestro project demonstrates orchestration across 30+ specialized agents. Client implementations use only the roles that the actual workflow and risk level justify.