Runs and costs
Every time an agent fires, it leaves a run behind — what triggered it, what happened, how long it took and what it cost. Open an agent to see its feed, newest first.
What the statuses mean
| Status | What happened |
|---|---|
| Queued | Accepted, waiting for a sandbox. |
| Running | The agent is working. |
| Succeeded | It finished on its own terms. |
| Failed | Something went wrong; the run says what. |
| Canceled | You stopped it. |
| Timed out | It stopped reporting and hit the time limit. |
| Skipped | A limit refused this trigger — the run never started. |
Succeeded means the run completed, not that it did what you hoped. An agent that decides there's nothing to do and stops is a success. If runs succeed but produce nothing you wanted, the prompt is the thing to change.
Skipped is not a failure either. It means a limit refused the trigger before any work started, so nothing ran and nothing was charged. It's a record that the trigger happened, kept so a gap in the feed doesn't look like an agent quietly breaking.
Cancelling a run
A live run has a Cancel beside it. Cancelling stops the work and tears down the sandbox, and the credential the run was using dies with it — an agent cannot keep acting after you've stopped it.
Anything the run already did stays done. A cancelled run that had already published a document leaves that document published; cancel stops the next step, it doesn't rewind the last one.
What a run costs
A finished run shows its credits. That figure covers the model tokens the agent spent and the time its sandbox was awake, and it settles as one line per run in your usage history — not as separate charges you'd have to recognise as belonging together.
Runs that never started — skipped — cost nothing. A run you cancel costs what it used up to that point, which is why cancelling early is worth doing.
In a shared project the project owner pays, with the run attributed to whoever created the agent, exactly like other shared work. See Pooled billing.
When nothing ran
If you expected a run and don't see one, the feed almost always says why.
A row marked "Skipped". A limit refused the trigger, and the row names which: too many runs going at once, or too many started in the last hour. Both scale with your plan. This is deliberate — a content type being republished in a loop, or an hourly schedule on a busy account, should slow down rather than spend everything you have. One skipped row stands in for a whole busy hour, so one row can represent many refused triggers.
No row at all. Then the trigger never fired, and it's one of:
- The agent is disabled — the toggle on the Agents page.
- The change wasn't one of the three content events. Editing a draft isn't an event; publishing, unpublishing and deleting are.
- The document belongs to a different project than the agent.
- A schedule hasn't come round yet. A missed slot is not made up later — see Creating an agent.
Runs that succeed while nothing happens. The likeliest cause is an agent with no Corte tools: it boots, thinks, and finishes, because every sandbox has its own file and shell tools — but it can't reach your content, media or workflows, so nothing in your account changes. The tool count under the access list says so before you create the agent and again when you edit it: see What it can touch.