The agent that keeps trying
Repeated tool calls increase cost without making progress. · PM Drills · 26 · 30 sec–2 min
The agent that keeps trying · 30 sec–2 min
Situation
An agent retries a failed supplier lookup twelve times.
The upstream service is unavailable. The user sees a spinner and cannot tell what is happening.
Think before scrolling
What product controls are missing?
Consider stopping rules, retry limits, progress visibility, and the fallback experience.
How a strong PM thinks
Bound the loop and define failure behavior.
Set limits on retries, elapsed time, tool use, and spend. Distinguish transient errors from conditions that require human or user action. Show a truthful state and preserve completed work.
A fixed retry policy with backoff may belong in software. The model should not invent success when tools fail. Evaluate the complete run, including its decision to stop.
Takeaway
Autonomy needs a budget and an exit.
A useful agent knows when further action is unlikely to help and makes the blocked state understandable.