Definition
Orchestration coordinates tasks, agents, tools, state, dependencies, budgets, failures, and lifecycle across a workflow. It decides what work is admitted, what can run now, which results unblock other work, how failures are handled, and when the overall run is complete.
The coordinating logic can be deterministic, model-driven, human-directed, or mixed. Durable orchestration persists state outside a model conversation so work can resume after a crash, avoid duplicate effects, and distinguish a completed task from one that merely stopped producing output.
Good orchestration makes ownership and join rules explicit. Parallel workers need non-overlapping authority or a conflict policy, bounded concurrency, input and output contracts, and a definition of when their results are safe to combine.
Distinguish it from nearby terms
Orchestration is the coordination function. A workflow is the process being coordinated. A scheduler mainly decides when work runs; an orchestrator also manages dependencies, state, failure, and lifecycle. Multi-agent only means that more than one agent participates.
Check your understanding
Five agents edit the same file in parallel, and the last write wins. The system has concurrency but lacks ownership and merge orchestration.