Skip to main content
Agent monitoring connects three kinds of evidence: what an agent did, how well it did it, and what its recorded work cost. Start with the question you need to answer, then follow a run back to its trace.

Traces

Send execution records, group them into sessions, and inspect model calls, tool calls, and errors.

Evals

Test fixed cases before a release and score recorded production sessions.

FinOps

Compare spend and tokens across agents, then check the rates behind a run’s cost.

​
Choose an investigation

​
Connect the evidence

A registered agent and a recorded run are separate records. Send telemetry with the agent’s own credentials and the intended workspace, then verify the trace arrived. Add session, visitor, and skill identifiers when the application knows them. A completed run can still produce a poor answer. Trace errors measure execution failures; eval scores measure the criteria your scorers define. Cost depends on the recorded token counts and the available rates. Read the trace, score, and price source together before deciding what to change.
Monitoring covers the work you send to Atlan. Missing spans, missing scores, and unavailable cost breakdowns do not establish that a run succeeded or cost nothing.

​
Start with one agent

  1. Register the agent and choose its workspace.
  2. Send and verify a trace.
  3. Inspect one run before relying on aggregate reports.
  4. Run a small eval and keep a failed production case in the dataset.
  5. Compare usage over the same workspace and period.