An orchestrator coordinates a workflow by choosing which component should run, what inputs it receives and when the result is ready for the next step. It can be ordinary application code, a workflow engine or another AI agent that plans and delegates.
A good orchestrator handles operational details such as task state, retries, timeouts, approvals and limits on cost or tool use. It may check the structure of each agent's output or send an uncertain result to a reviewer before continuing. Those checks make failures easier to diagnose.
Allowing a model to control every transition can provide flexibility but also produce loops, unnecessary delegation or unexpected actions. A fixed sequence in code is usually easier to test when the steps and success conditions are known. Delegation itself is not proof of correctness.
Choose the simplest coordination method that solves the task. Define clear completion conditions, enforce permissions outside the model and keep an event trace showing what was assigned, returned and approved.