When Should an AI Workflow Stop Instead of Retry?

A retry can be useful when a temporary failure is likely. But if the workflow’s state has changed—or an action may already have happened—repeating it can create a second problem.

What should trigger a pause, a retry, or a handoff to a person? For example, would you base that choice on whether the action is reversible, whether its completion can be verified, or whether the system still has the context it needs? I’m especially interested in what makes the next owner’s responsibility clear when the automated route stops.