Run controls, monitoring and reports

Controls while a run is working

A run is a live process, and you can intervene.

• Pause stops a run at its next safe boundary, so nothing is cut off mid-step. A paused run holds no resources and nothing happens to it until you decide.
• Resume carries on from where it stopped. The run picks up using its recorded steps.
• Retry starts a new run with the same input after a run has ended, whether it was done, failed or cancelled. Use it after fixing a cause.
• Cancel ends it for good. Use it the moment you realise something is wrong.

Which buttons you see depends on the run's status: a running run offers Pause and Cancel, a paused run offers Resume and Cancel, and a finished run offers Retry. Test and replay runs are repeated with Replay instead. If you press Pause on a running run it says "Pausing…" until it reaches a safe point, and you can choose Keep going to withdraw the request.

Monitor in layers

One run. Open the run page for the live trace. The statuses are Queued, Running, Paused, Waiting for your approval, Done, Failed and Cancelled.

All runs. The Runs page lists everything newest first. Filter by Agent, Result, How it started and When. This is where you look for "all failures this week" or "everything that came from the webhook".

Trends. Reports shows how your agents are doing over seven, thirty or ninety days: runs, Success rate, Typical run time, Cost, Runs per day, Approvals, Most used tools and Recent failures. An agent's own page has an Activity section and a report for just that agent.

Money. Usage shows tokens and cost per agent against the monthly budget.

What to look for

• Falling success rate: something changed in an input, a tool or a model. Open Recent failures and read the traces.
• Rising cost or time: the agent is taking more turns or reading more. Check the typical run.
• Many approvals denied: the instructions are producing requests you do not want.
• A tool used far more than expected: it may be looping.
• Approvals waiting for days: either nobody is watching or the agent asks too often.

A weekly review habit

Once a week, for each agent that runs without you: open Reports, scan Recent failures, glance at Usage, and read one successful run end to end. Fifteen minutes finds most problems early.

Try it yourself

1. Start a run of an agent and press Pause while it works. Then choose Resume.
2. Open a finished run and choose Retry. Compare the two runs.
3. Open Runs and filter to failed results. Open Reports and choose 30 days.

Recap

• You can Pause, Resume, Retry and Cancel a run, depending on its status.
• Monitor one run in its trace, all runs in Runs, trends in Reports and money in Usage.
• Look for falling success, rising cost, many denials and unexpected tool use, and review weekly.