Reading a run
What every part of the run page is, and which parts you can trust.
The run page is where the platform has to be honest. Everything on it was recorded while the agent worked — nothing is reconstructed afterwards, and a run that has not reached a phase does not show one.
It updates itself while the run is live, and follows the tail of the transcript until you scroll up to look at something.
The tiles
| Tile | What it is |
|---|---|
| Duration | Wall-clock from start. Says still running while it is. |
| Score | Criteria passed over criteria judged. — when there was no rubric. |
| Tokens | What the agent reported using. |
| Cost | The same, in money — spent on your own provider account. |
Instructions
The task exactly as the agent received it, with the machine shape and any files the task carried. If the task has since been deleted, it says so rather than showing an empty brief.
Output
The agent’s final answer, rendered as Markdown. While the run is live this is empty and says so — the transcript underneath is the live part.
This is the agent’s account of its own work. It is not evidence. That distinction is the whole reason the next two sections exist.
Rubric
One line per criterion, in the order you wrote them, each with a tick or a cross and the evaluator’s note underneath. A note cites the concrete thing that was found — a path, a command, a line of output.
A criterion still shows a neutral mark and Not judged yet while judging is in progress. Underneath, when there is one, is the evaluator’s summary of what the agent actually did.
If no rubric was chosen, the section says so: the verdict is whether the agent’s own process succeeded.
The agent’s transcript
Everything the agent thought, called and produced, in order, with a count in the heading. Tool calls show their arguments, and what came back is one click away — so a claim in the final answer can be checked against the call that supposedly produced it.
For a catalog agent this is parsed from the CLI’s own structured output. For
your own agent it is what the SDK captured: your framework’s spans, plus
anything you reported through ctx.
Evaluation
A separate record, deliberately. Everything from the moment judging starts is our evaluator rather than the agent — its reasoning, the tools it reached for, and what each returned, kept in the shape the tool printed it.
Read together with the agent’s transcript they would be confusing, and the point of the page is to see what the agent did. Read separately, this section is the audit trail for every verdict above it.
When a run did not complete
An alarm panel at the top carries the reason in plain words: a missing credential, a required setup step that failed, a timeout, a crash. The transcript above the failure is still there, and is usually where the reason becomes obvious.