Traditional monitoring tells you whether a service responded. Agent observability must also explain what the system understood, which context it used, why it chose an action and what happened next.
What should an agent trace contain?
Capture the decision chain without storing unnecessary sensitive content.
A useful trace links the initiating event, retrieved context, model decision, policy evaluation, tool call and final business outcome under one correlation identifier.
- Input and actor reference
- Model and prompt version
- Retrieved sources
- Tool request and validated response
- Policy decisions
- Cost, latency and outcome
Measure the operation, not only the model
Accuracy is incomplete if the process becomes slower, more expensive or harder to recover.
Track completion rate, exception rate, review time, cost per successful outcome and the share of actions reversed by people. These connect technical behaviour to operating value.
trace_id | decision | policy | tool | outcome | latency | costDesign for investigation
A team should be able to reconstruct a consequential action without replaying private data through a model.
Version prompts, policies and tool schemas. Record references to protected data rather than copying it into every log. Set retention according to operational and regulatory needs.