An agent is slow, and the team must distinguish model latency from retrieval and external tool delays. Which approach best fits?

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