Run Inspector and agent activity
Use agent activity to follow work. Use Run Inspector when you need to explain which model or tool handled a particular turn.
Follow background work
Open the chat’s Agents pop-out to see active workers, their current tool, and completed results. Ask “Is the researcher still working?” or “Check workers” for status. Use Stop to cancel work; cancellation follows the owned child jobs. A job is not complete just because its parent finished speaking.
For scheduled runs, open Tasks → Ledger. Their output stays in the ledger. For an interrupted supported worker, use its Review & resume controls; see Resumable agent jobs.
Inspect each agent’s work
Open Agents beside the chat composer, then expand Progress, actions & conclusion under a task. Everyone can inspect work from their own conversations; this view does not require owner or administrator access.
Each task shows its assignment, current activity, reported milestones, recorded tool inputs and results, and the agent’s conclusion or failure report. For a four-agent job, expand each agent separately to compare its work and conclusion. The view refreshes while open. An agent that has not finished has no final conclusion yet; a tool interrupted without a recorded result is marked outcome not confirmed. A conclusion is the agent’s report, with observed tool results shown separately for checking its claims.
Use Download work record to keep a JSON copy of the displayed evidence. Completed work reloads from the original conversation, including worker reports that do not appear as separate chat replies. It follows conversation retention and deletion: older project conversations remain in Progress, and clearing general chat removes its conversation record. Older tasks can have fewer details because only evidence that was recorded at the time is available.
Long tool inputs and results are shortened, with a notice; credential values are redacted. The detail view displays at most 256 actions and the last 200 progress updates and identifies omitted entries. It shows observable work and reported conclusions, not the model’s private reasoning.
Inspect a turn
Run Inspector is an owner/admin diagnostic. Open Run Inspector from the sidebar, or its tile in the mobile menu. Select the relevant run and compare:
- the input source and matched skills;
- requested and resolved provider, model, and reasoning effort;
- tools offered to the model and actual tool-call results;
- separate local records and provider-reported evidence;
- failures and any actions that already completed before a later error.
An unsupported or unavailable model is a setup problem, not proof that a tool failed. A tool that completed before a provider error may already have changed data. Check its result before retrying the whole request.
Why did a tool not run?
Ask an administrator-enabled assistant “Why isn’t my tool being called?” OE’s diagnostic checks the capability gates, including manifest/handler registration, agent assignment, account permissions, routing, and voice access. Correct the reported gate, then retry one small request.
For a custom skill, confirm the assigned agent in Settings → Skills → Custom Skills. For MCP, confirm Assigned to on its server card. A skill intended for voice use must declare that capability; see Building custom skills.
Approval cards and errors
When OE stages a consequential action, review the current Approve / Cancel card. Cards remain associated with their request after reload; a newer request can replace an older pending action. Approving the current card authorizes the action shown there.
Short notices also report recovery such as token renewal or unsupported provider capabilities. Inspect the run when you need details. Estimated token counts are labeled as estimates; missing provider usage is not an exact zero.