Every real pipeline has an orchestration layer whether or not anybody built one deliberately. If nobody did, it is a person with a mental checklist, and it stops working the week they are away.
What it has to handle is mostly unglamorous: ordering, retries with a limit, partial failure, and the escalation points where a human decision is required. Retries without a limit are how an unattended job spends a budget overnight.
Keep the orchestration deterministic even when the steps are agentic. A model choosing what to do inside a step is useful; a model choosing whether the step runs at all is a system nobody can predict or audit.
Should the orchestration layer itself be agentic?
No. Keep judgement inside steps and keep the control flow deterministic, or you get a system whose behaviour cannot be predicted, and which nobody can reproduce or audit afterwards.
What is the most common orchestration failure?
Unbounded retries. An unattended job that keeps trying is how a credit ceiling gets discovered the morning after it should have stopped something.