The Trace Design Prompt

Design observability for this agent.

Agent: [DESCRIBE]
Who debugs it: [DESCRIBE]
Typical debugging question: [E.G., "why did it do that?"]

Produce:
1. What to log at each step of the agent loop
2. Trace structure: how a run is represented so it can be replayed
3. What to capture per tool call: arguments, result, latency,
   error, retry count
4. Token and cost accounting per step
5. What to capture about model decisions: which tool, why (if
   available), what alternatives existed
6. Correlation: how to tie a run to a user, a session, a request
7. Sampling: what to keep for every run vs only for failures
8. Retention: how long, and what regulation applies
9. What must be redacted before logging

Then:
10. Given this trace, could someone answer "why did the agent do X
    on step 12?" three weeks later? If not, what is missing.
11. What alert should fire automatically, and on what signal