Agent tracing

Agent tracing is recording every step of an AI agent's run in order (each message, model call, tool call, sub-agent and guardrail verdict), so the run can be read back later to debug it or audit it.

What a trace holds

A trace is one run from the first message to the last. The terms come from distributed tracing: a trace is a tree of spans, each one a step with a start and an end.

  • Messages. The user's request, the system prompt, each model response.
  • Tool calls. The name, the arguments, the result, the time it took.
  • Sub-agents. An agent started by another, with its own steps nested under the call.
  • Verdicts. Which policies were checked at a step, which fired, what was done.

Traces are the raw material of AI observability.

Why you need it

  • Debugging. The final answer rarely shows why a run went wrong. The cause is often many steps back: a bad search result, a tool error the model worked around. A rerun may take another path, so the trace is the only record.
  • Auditing. After an incident you have to show what the agent read, what it did and for whom. See audit logs.

Traces hold personal data and secrets. Redact them and set a retention period.

With White Circle

Analyze has a session inspector that shows every event of a session and which policies fired. Logging is opt-in, with control over redaction and retention.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL