An agent repeatedly calls the same tool without making progress. Which approach best fits?

1. Measure token use, tool steps, retries, and end-to-end cost per task.
2. Trace the workflow stages and record per-step latency.
3. Use bounded retries, timeouts, and appropriate circuit-breaking behavior.
4. Add explicit iteration limits, progress checks, and a controlled failure path.