Reading a run
Read what one run did, find where it went wrong, and open any step's details and raw log.
Overview
When you click a run in the Agent Runs list, it opens in the run view, the middle pane. Click a step in the run view's trace and it opens in the step inspector, the right pane. The run list stays beside them, so you never lose your place.
The run view shows, from top to bottom:
- The run header. What the run was, how it ended, how long it took and what it cost.
- The summary card, with the run's issues and a likely fix, beside Where it went.
- The trace. Every step of the run on a time axis.
- Run facts. IDs, models, tools and log storage, folded away until you open them.
On a screen narrower than 1200px, the run view sits below the run list, and the step inspector slides in from the right when you click a step. Press Esc or click × to close it.
The run header
The header shows:
- The title. What the run was about. A run with no agent, such as a dream cycle or a mailbox scan, is titled by its source and carries a badge such as Dream or Scan.
- The owner line. For a Personal Assistant run, the owner's name and email. For other runs with a mailbox, the mailbox address.
- The status. A pill such as Completed, Completed with issues, Failed, Waiting, Escalated, Stopped by a policy or In progress. The pill always has words and an icon, never colour alone. It matches the run's dot in the list.
- The reason. When a run did not end normally, a small chip names why, for example "Max Tool Calls Reached".
- The agent and what started the run, such as "Portfolio run".
- The start time, the duration and the cost. The cost includes any sub-runs the run started.
- Counts: model calls, tool calls, and how many tool calls failed. A tool call held for review is not counted as a failure.
On the right of the header:
- Newer and Older move to the next run in the list. A position line such as "7 of 20" shows where this run sits in the list. It reads Not in this filter when the run does not match the list's current filters.
- Copy link copies a short link to this run.
Agent-to-agent messages
When a run was a message from one agent to another, a strip under the header shows who sent it and who received it, the hop count, and when it was delivered. It can link to View A2A thread and to View recipient run (best match). The recipient run is a best guess: the recipient agent's nearest run within an hour of delivery.
The summary card
The AI summary card tells you what the run did, in plain words.
- If the agent wrote its own summary, the card shows it, labelled "Agent's own summary".
- Click Summarize to have AI write a summary of the run's steps. The summary is saved with the run, so the next person who opens it sees the same summary. The card names the model that wrote it and when.
- Click Regenerate to write a fresh summary that replaces the saved one.
- A skipped run shows why it was skipped instead, and offers neither button.
Summarize is not available while the run is still going ("Available when the run finishes"), or when the run has no step log ("No step log to summarize"). In an impersonation or partner session, the buttons are hidden, and saved summaries still show.
If a summary cannot be written, a message shows inside the card, under the buttons. The summary you already had stays in place.
Issues
Under the summary, each issue the run hit is one row, with an icon and a colour:
- Red: an error, such as a tool call that failed.
- Violet: something held or waiting for a person, such as an approval.
- Amber: a warning, such as a reply that was delivered in a degraded form, or a run that is queued to retry.
When an issue belongs to a step, its row has a Step n › link. Click it to select that step in the trace and open it in the step inspector.
Likely fix
When the platform recognises the problem, a blue likely fix callout says what probably went wrong and what to do about it. It can link to the step that caused it, and it can offer buttons that take you where to fix it. A button you cannot use shows as plain text.
Parent link
A run that another run started links back to it. A sub-run reads "Part of [parent run] (sub-run)". A run started by another agent's message reads "Started by [parent run] (hand-off)". Click the link to open the parent run with the step that started this run already selected.
Where it went
The Where it went card shows two bars:
- Time by kind. How the run's time split between thinking, tools and memory. The rest is "Waiting and overhead".
- Cost by phase. Which part of the run cost the most, largest first. Sub-runs are their own phases.
A cost that no step accounts for shows on its own row, such as "Not attributed to a step". A note shows the cost of the pre-execution classifier when it ran, because it is billed separately from the run.
The trace
The trace is the run's runbook: every step the run took, in order, drawn on a time axis. It replaces a long list of log entries with one picture of the run.
The trace header reads like "Trace · 68 steps in 5 phases, 3 sub-runs", with counts of model calls, tool calls and tool searches. These counts come from the step log, so they can differ from the header's counts.
Phases and steps
- Phase rows group steps, such as Start, Work and Wrap up. A sub-run is its own phase, with a Sub-run badge. A phase row shows its number of steps, its duration and its cost, and a coloured dot for each step with a problem.
- Step rows show the step number, an icon for its kind, a title such as "Searched the web for supplier pricing", and a status icon when the step did not go well. Each step's bar shows when it started and how long it took.
- Problem rows are tinted red, violet or amber. The selected step is blue.
The legend names each kind of step (Thinking, Tool, Memory, Lifecycle, Safety check, Hand-off, Error) and each state (Instant event, Running, Recovered, Failed, Waiting for approval). Every state has its own shape or pattern as well as a colour.
Folding
Click a phase row to fold or unfold it. Collapse all and Expand all fold or unfold every phase.
When a run has more than 20 steps, sub-runs with no problems start folded. A phase that holds the selected step, a running step or a problem never starts folded. Your folding stays as you set it while the run refreshes.
A very long run with more than 300 steps in one phase groups its steps by turn, "Turn 1", "Turn 2" and so on.
Size by
Size by: Time | Cost switches the bars between two views:
- Time draws each step at its time from the start of the run.
- Cost lays the model calls end to end by what they cost, so the costliest steps are the widest. It is disabled when no step has a cost.
The page remembers your choice in this browser.
Selecting a step
Click a step row or its bar to open the step in the step inspector. The first time you open a run, the page selects its first failed step, or else its first held step, or else its first warning, or else step 1.
With the trace focused:
| Key | Action |
|---|---|
| ↑ / ↓ | Previous / next step |
| ← | On a step, go to its phase. On a phase, fold it |
| → | Unfold a phase |
| Enter or Space | Fold or unfold a phase |
| Home / End | First / last step |
The step inspector
The step inspector shows the selected step's detail.
Its header shows the step's kind and status, its title, its clock time, its time from the start of the run, its duration and its cost, and "Step n of N". The status always has words: "Failed", "Waiting for approval", "Recovered" or "Running".
Use the ˄ and ˅ buttons, or ↑ and ↓ with focus in the header, to move to the previous or next step.
Tabs by step kind
The tabs depend on what kind of step it is. Press 1 to 5 to switch tabs.
| Step kind | Tabs |
|---|---|
| A model call (thinking) | Overview · Prompt · Response · Tokens · Raw log |
| A tool call | Overview · Input · Output · Raw log |
| Any other step (memory, lifecycle, safety check, hand-off, error) | Overview · Details · Raw log |
The inspector remembers the tab you chose. When you move to a step of another kind, it shows the matching tab. For example, Output on a tool call becomes Response on a model call.
| Tab | What it shows |
|---|---|
| Overview | What the step was, its phase, when it started and how long it took. A problem note says what went wrong. A model call shows the question it was asked, what it decided and the tools it asked for. A tool call shows its result: OK, Failed, "Failed, then OK after retry", or "Held for review". A held step can link to the place where you review it. |
| Prompt | What was sent to the model: the system instructions, the agent's memory, the user message, tool results and the tools the model could use. Click a tool to see its full schema. Switch between Preview and JSON. |
| Response | What the model sent back: the text, the tools it asked to call, and why it stopped. A "hit token cap" chip shows when the reply ran out of room. |
| Tokens | The step's tokens split into new input, cache write, cache read and response, with the step's cost, its prompt cache hit and its share of the run's cost. |
| Input | The parameters the tool was called with. Switch between Table and JSON. |
| Output | What the tool returned, or the error it returned. Switch between Table and JSON. |
| Details | The detail for other kinds of step, such as a memory update, a Safety Gateway decision, an agent-to-agent message or an error message. |
| Raw log | The step's entries in the audit log. |
A very large prompt, response, input or output was kept as a placeholder in the audit log. Its tab shows a Truncated at 1 MB chip.
Raw log
The Raw log tab lists every audit-log entry behind the step, and says whether the log was read from your own storage (BYOS) or from Outermind storage. Each entry shows its type, its time to the millisecond, its level and its size.
- Entries start collapsed. Use Expand all and Collapse to open and close them.
- Each entry has its own copy button. A very large entry has Download entry instead.
To keep the whole run's log, open Run facts and click Download full log (JSON).
Sharing a run or a step
- Copy link in the run header copies a short link to the run.
- Copy link in the step inspector (the link icon) copies a short link to the run with that step and tab selected.
Anyone who can see the run can open the link from any hub. The link opens the run and the step, not the list filters you had.
The step inspector also has Copy this step's log entries, which copies the step's raw entries as JSON. A very large step offers Download step JSON instead.
Run facts
Run facts is a folded card at the bottom of the run view. Click it to open it. It holds:
- IDs and source: the execution ID with a copy button, the agent, the owner, what started the run and how it ended.
- Configuration: the executor type, work type, prompt tier and models used.
- Tokens and tool calls: request and response tokens, the prompt cache hit, the tool calls that succeeded and failed, the tool-result cache and a count of calls by tool name.
- Pre-execution classifier: the routing decision, its confidence, latency, model and cost.
- Sub-runs: one row per sub-run, with its agent, status, duration and cost, and a link to open it.
- Current log storage: where your tenant's logs are stored now.
At the bottom of the card:
- Download full log (JSON) saves every log entry for the run as a file.
- Copy response JSON copies everything the page loaded for the run.
Live runs
A run that is still going updates on its own.
- The status pill shows In progress, and the duration counts up.
- New steps appear in the trace about every 12 seconds.
- A step that is still running is drawn up to now with a striped end and reads "Running". Its Response or Output tab says it is waiting.
- The step you selected stays selected while the run refreshes.
- Summarize turns on once the run finishes.
Privacy-locked runs
A Personal Assistant's privacy policy can hide the details of its runs from you. A privacy-locked run shows only:
- the time, status, agent, the owner's name, what started the run, the duration, the cost and the counts;
- the card "Details are hidden by your PA privacy policy.";
- Run facts, with only those same fields.
It has no summary, issues, Where it went, trace, step inspector or raw log. A sub-run that is privacy-locked shows in the trace as a locked phase with its duration and cost only.
A run you are not allowed to see at all reads "This run doesn't exist or you can't see it."
When the step log is missing
The header still shows the run's counts and cost. The trace card says why there are no steps:
- "Step logs for this run are no longer available." The logs are past your retention period.
- "No step log was found for this run."
- "No step log was found in this tenant's own log storage for this run." Your tenant stores its own logs (BYOS).
- "No step log was found, and the retention policy couldn't be checked." Click Retry.
Troubleshooting a run
- Read the header's status and reason.
- Read the summary card's issues and the likely fix.
- Click Step n › on an issue to open the step that caused it.
- In the step inspector, read Overview, then Input and Output for a tool call, or Prompt and Response for a model call.
- Open Raw log to read the exact audit-log entries.
- Use Copy link to share the step with someone who can help.