Skip to main content
The Traces page lists the turns of your own AI tool sessions and the runs of every agent you can open, so you can pick a turn or run from any period and read its trace beside the list. Use it when you want to find what to investigate rather than read one run you already have. To compare agents across the fleet first, start from Compare agent usage; to read one run’s trace in depth, see Inspect recorded run.

​
Before you begin

Before you start, make sure you have:
  • Access: Agent runs appear for every agent you can open. Your AI tool sessions are your own.
  • Turns to read: Turns appear once an AI tool sends its first trace; to capture Claude Code and Codex sessions (Atlan shows Codex as ChatGPT), see Capture coding-agent traces with aireg.
1

Choose view

In the sidebar, select Traces, below Agents, and choose the period from the date range picker in the page header. Two views make up the page, as separate lists, newest first:A count beside each tab covers the whole period, so picking a filter never moves it. A row’s title is its opening prompt, and the dot beside it gives its status; a failed row also carries a Failed chip, and one that finished with errors names how many. Model shows the model that answered over the AI tool that ran it. Work shows how long the turn or run took, with its steps and tokens on the line under it. Once you pick one agent, the Agents view trades its Agent column for Started by.
2

Narrow the list

  1. On AI tools, pick a Tool. On Agents, filter by Agent or by Workspace.
  2. On either view, filter by Status: Completed, Failed, or Running. Running covers a run still recording activity; a run abandoned days ago reads Completed.
  3. Search the list: a search matches the opening prompt, agent, tool, or model of the loaded rows.
  4. When nothing matches, Clear filters clears the search and the Status and Verdict filters.
Filters and search narrow the loaded rows only; page back through the footer for older runs. The footer reports what matches, such as “12 of the 40 loaded runs match”, and an empty list says “Older runs may match. The figures above still cover the whole scope.”
3

Read what changed

What changed above the list shows the period’s five figures: the count of turns or runs, the share that finished without a recorded error, p90 duration, tokens per turn or run, and recorded cost per turn or run. One turn is one trace an AI tool sent; one run is one end-to-end execution attributed to an agent. A line under the heading says what the figures stand on, such as “Based on 212 turns from Claude Code and ChatGPT.”Each figure compares with the previous period when both periods had at least ten runs; otherwise it stands alone and the figures say so. Error-free turns (on Agents, Error-free runs) does not measure answer quality — a run that finished without a recorded error can still fail a quality rubric. Across agents, p90 duration shows a dash until you pick one agent.
4

Read scores

When Evals is enabled for your organization, picking one agent on the Agents view adds that agent’s scores:
  • The Scores strip shows one figure per scorer: its mean score when it answers with a number, its share of passes when it has a pass line, or its commonest answer when it answers with words. Each figure shows its number of verdicts, plus how many fell under the pass line for a mean or how many failed for a share of passes. The heading says which runs the figures stand on, such as “Over the 40 loaded runs: 36 carry a verdict.”
  • The strip also shows which scorers run automatically, and Manage scorers opens the scorers page. When the agent has no scorer set to run automatically, the strip says so and offers Open scorers. When scorers run automatically but none has judged the loaded runs yet, no strip appears.
  • The run list gains one column per scorer, at most three, with that scorer’s verdict for each run and a dash where there is none yet. Across agents, a Flagged column counts each run’s verdicts under a scorer’s pass line instead.
  • The Verdict filter offers Flagged, Passed, and Not scored yet, and Show flagged runs on the strip filters the list to flagged runs. A verdict without a pass line, or one whose label the scorer never rated, counts as neither flagged nor passed.
5

Open a run

  1. Select a row to open its run in a sheet beside the list. The sheet’s header shows who ran it — the agent, or the AI tool — its status, when it started, who started it, and the start of its trace ID. The title is the run’s opening prompt.
  2. Read the run in the sheet’s views: Highlights for the opening prompt, key decisions, errors, and closing response; Full trace for every recorded step; Waterfall for the same steps on one time axis; Scores for the verdicts the run’s scorers left. Highlights is the default.
  3. Move along the list with Previous run and Next run, or the up and down arrow keys; the sheet keeps the open row in view.
  4. Select Open full page to read the run on its own page, or close the sheet to return to the list.
6

Verify what covers what

  • The figures cover the whole scope for the period; the list covers its loaded rows. A filter that empties the list leaves the figures whole.
  • A run is filed under the account that sent it. A run started from a terminal is listed under that account rather than under an agent; see How trace attribution works.
  • When some agents’ runs could not be loaded, the page says which and shows the others.

​
Troubleshooting

​
Next steps

​
See also