AI agent orchestration decides what happens around model calls: which tool may run, when a result is checked, whether another agent should take over, and when the task must stop or ask a person for help. It can be implemented in code, a workflow platform or an agent framework.
Think of it as traffic control around the model. In a refund workflow, orchestration might let the agent look up an order, require a person to approve any refund above a set amount, cap the run at ten steps and write every action to a log. The model suggests the next move; the orchestration layer decides whether that move is allowed. Many AI agent frameworks provide this layer ready-made, and a multi-agent system needs it even more. Approval steps like this are often called human-in-the-loop design.
Good orchestration makes the process observable and bounded. Define allowed actions, timeouts, retry limits and clear completion criteria rather than letting the agent continue indefinitely. Related: Workflow, Tool Calling, Error Handling.