PRODUCTION SCENARIO
An autonomous data-cleaning agent on Claude Sonnet 5 works through batches of financial ledgers. On a row with a malformed date it calls parse_date with the same arguments again and again. Every one of those calls returns in about 40 ms with HTTP 200 and the same unparseable payload, so nothing in the transport layer looks wrong, and the loop runs until the task exhausts its tokens and the rest of the batch is left unprocessed. Operations needs the agent to notice the dead end while it is happening, change approach on that row, and carry on through the remaining ledgers.
Which control stops the repetition while the run is still usable?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Hash the tool name and argument set on each call, returning an error on the third repeat
The calls succeed in 40 ms and return the same payload every time, so the only thing separating the loop from useful work is that the name and arguments repeat; hashing them lets the harness hand back an error the model can act on, on the row where it happens, while the batch keeps moving. Exponential backoff is the right answer when the repeats come from a failing dependency, but nothing here is failing, so a retry policy has no fault to wait out. A wall-clock deadline is the control for a call that hangs, and these return immediately. An advisory task budget is what you reach for when spend per ledger is the constraint; it shapes how the agent wraps up rather than breaking the loop, which can still consume the whole budget.