An agent's cost increases after a prompt change despite stable user traffic. Which approach best fits?

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