4.7 — Observability (traces)

🟢 In plain words — observability is the agent’s black box. It records observable events; it does not necessarily expose the model’s private reasoning.

Useful traces may contain requests and responses according to your data policy, tool calls and arguments, statuses, errors, duration, token usage, estimated cost, approvals and final outcome.

When a run fails, inspect the facts: input, selected tool, arguments, result, approval and final decision. Then correct instructions, data, tool, permissions or workflow. A trace is evidence, but it does not prove compliance by itself; policies, controls, retention and access rules are also required.

✅ Summary

• Trace observable inputs/outputs, tools, errors, duration and consumption.
• Do not confuse traces with hidden chain-of-thought.
• Minimise sensitive data and control access and retention.