AI traces
Staff-only inspection of recorded agent runs, tool calls, decisions and captured context.
Console access
Console → AI traces is available to platform superadmins and admins. Organisation owners and administrators do not gain Console access through their workspace role. Support staff, sales staff and customers cannot open this page or read its API. Settings is the last item in the Console menu for staff who can access it.
Find a run
The page opens the most recent recorded run. Search recorded input previews or session IDs, or filter by agent, status and tool. Clear filters returns to the recent list. Selecting a run keeps your filters and current page in the address, so you can share that address with another authorised staff member. Refresh reads the event journal again.
This view scans up to the most recent 500 sessions on the API host and shows 10 matching runs per page, across at most 50 pages. Previous and Next keep your search and agent/status/tool filters and select the first run on the new page. Changing a filter starts at page one and clears the run and context selection. Enter a number in Page and select Go or press Enter to jump directly to that page. Choose a whole number from 1 to the displayed page count. A jump keeps your filters and selects the first run on the new page. The page number and range show where you are in the matching results. With one page or no matching runs, the page field and Go button are disabled.
Search and filters apply before pagination. The 500-session bound is a scan limit, not a storage guarantee or retention rule. Older sessions and runs on another host are outside this view; an empty result does not establish that no such run exists in older history.
Read the health figures
The last-seven-days figures describe all runs found in the scanned sessions. Filtering the list does not change them. Runs by day uses UTC dates; run timestamps use Australia/Sydney time.
Failure rate and median/p95 latency use completed and failed runs. Other includes open runs, refusals and runs waiting for approval. Tokens are the input and output usage recorded in the journal. A tool can fail during a run that later completes; the tool-error list keeps those failures visible.
Inspect the timeline
Expand a timeline row to inspect the values the journal recorded:
- Model: model name, tokens, cached tokens and stop reason, when recorded.
- Tool: policy, arguments, output and error.
- Guardrail: action, provider, model, answer, probability and reason, when recorded.
- Approval: reviewed arguments, approval ID and recorded decision. A pending approval remains visible even if the agent run has completed.
- Subagent: the recorded start and end of the delegated work.
An open span has no recorded end. The duration shown for a run ends at its last recorded event; it is not a timer. The page is an inspection surface: it does not approve actions, retry tools or start model runs.
The Input panel starts collapsed to keep long prompts compact. Expand it to read the full recorded prompt. Selecting another run starts its input collapsed. Raw events exposes the recorded event payloads, clipped by the trace contract where necessary. Expand Session ID and eval capture to copy the ID or the command for capturing an eval case on the API host.
Captured context
Show captured context appears only where a model-call event points to captured messages. Context capture is optional and retained separately from the journal. When capture was off, the adapter did not supply messages, or retention removed the file, the page says that context is unavailable. A pointer in an old event does not guarantee the messages still exist.
Inputs, tool arguments, results and context can contain information from other workspaces. Their API requires the same completed sign-in and two-step verification as the Console page, and a platform admin role, on every read.