# What is Agent Orchestration?

Also called agent coordination.

Agent orchestration is the layer that decides which agent or step runs, in what order, with what inputs, and what happens when something fails. It covers routing, scheduling, retries, state passing, and concurrency limits. Orchestration is ordinary software rather than model output, which is what makes multi-step agent systems observable and recoverable.

An orchestrator holds the parts of a run that must not be left to a model. It records which step is active, persists intermediate results so a crash does not lose the work, enforces timeouts, retries transient failures, and applies limits on how many steps or how much spend a single run may consume. Agents make decisions; the orchestrator makes those decisions durable.

Routing can be decided in code or by a model. Code routing uses conditions on structured outputs and is predictable. Model routing lets an agent name the next step, which handles cases nobody enumerated but can select a step that does not exist or loop. Many systems use code for the outer shape and model choice only inside a bounded set of options.

Long-running work makes durability the central concern. A run that spans minutes or days will hit an expired token, a rate limit, or a restarted process. Durable execution engines address this by checkpointing every completed step so a resumed run replays from the last good state instead of starting over, and so partial side effects are not repeated.

Orchestration is sometimes mistaken for the supervisor agent pattern. The supervisor is an agent that reasons about delegation. The orchestrator is infrastructure that executes and records the resulting plan. A system can have one, both, or neither. When people say an agent framework is unreliable, the missing piece is usually orchestration rather than model quality.

## Key points

- Orchestration is code, not model output, and stays predictable.
- Handles routing, retries, timeouts, state persistence, and step limits.
- Checkpointing lets long runs resume instead of restarting.
- Distinct from a supervisor agent, which reasons rather than executes.

## In practice

A five-step research run reaches step four and the search API returns a rate limit error. The orchestrator does not lose the first three steps. It waits, retries twice, and on the third failure marks the run as needing attention with everything completed so far intact. When the run resumes an hour later, it picks up at step four rather than the beginning.

## Related terms

- [Multi-Agent System](/en/glossary/multi-agent-system)
- [Agentic Workflow](/en/glossary/agentic-workflow)
- [Supervisor Agent](/en/glossary/supervisor-agent)
- [Agent State](/en/glossary/agent-state)
- [Handoff](/en/glossary/handoff)

[Back to the AI Glossary](/en/glossary)