What should be decided before an AI task runs repeatedly?
Before an AI task runs repeatedly, decide what condition it checks and what response it's allowed to make. Add limits and stopping rules, and name an owner for the unresolved work. Check that canceling it stops future actions.
A repeated task carries a different responsibility from a one time request. The Power Shift discussion of more autonomous work is a good reason to decide up front when repetition is useful and when it should end.
Every repeating task needs an off switch
Set the frequency from the real need and the relevant service's constraints. More activity isn't a measure of usefulness. Put a limit on messages, attempts or costs where they apply, and write stopping rules for errors and unavailable access, and for an outcome someone else already completed.
Put one person in charge of what's left unresolved. They should know what the workflow is waiting for and when it checks next. They should also know what happens if the condition never arrives.
I'd test cancellation before launch and confirm it stops future actions. A recurring workflow should be able to explain both why it runs and why it will stop.