Your question is Observability for Agent Decisions. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
How would you architect an observability framework to audit agent decision-making in a production system?
Explain how you would capture agent inputs, plans, tool calls, intermediate state, model outputs, policy checks, and final actions without exposing sensitive data. Cover online tracing, asynchronous analysis, decision reconstruction, evaluation, alerting, retention, access control, and response to anomalous or unsafe behavior. Clarify how the design handles non-determinism, retries, partial traces, prompt or model changes, training-serving skew, and failures in external tools.