What is Agent orchestration?

Agent orchestration is the layer that coordinates how an AI system gets work done when a single model call is not enough. In its simplest form it sequences the steps of one agent's loop; at the more ambitious end it directs a multi-agent system, deciding which specialized agent handles which subtask, in what order, and how their outputs combine. An orchestrator handles control flow (sequential, parallel, or conditional branching), passes context and intermediate results between steps, retries or reroutes on failure, and enforces budgets so a runaway loop does not spiral. A recurring pattern is a planner or supervisor agent that decomposes a request and delegates to worker agents, then synthesizes their results. The Model Context Protocol matters here because orchestration is mostly about tool access and shared context: MCP gives every agent in the system a uniform way to reach the same tools, and a shared memory store keeps their understanding consistent so one agent's discovery is available to the next. Without coordinated memory and tool access, orchestrated agents drift, duplicate work, or contradict each other across steps.