all-agents docs
Runs and observability

Stopping a run

The emergency brake on the run screen, what it stops, what it cannot stop, and how to stop a whole workflow at once.

Every run that is still going has a red Stop button on its run screen. It is offered only while the run can actually be stopped, because a button that does nothing is worse than no button when somebody is in a hurry.

What the dialog asks

Abort this run? It stops recording anything further right away. Work already sent to another system has been sent.

Under that is a checkbox, ticked by default:

Also stop this workflow's other running runs and pause the workflow

It is ticked on purpose. What an owner has usually just watched is not one unlucky run, it is a flaw in the workflow itself: a hardcoded address with a typo, a filter that matches everything. Stopping this one run leaves the trigger free to produce the next one, and cancelling them by hand is a race nobody wins. So the default action stops every live run of the workflow and pauses the workflow, in one act.

Untick it if you really do mean this run only.

Then it asks what went wrong

Once the run is stopped, the dialog asks What went wrong? with a text box. That step comes second because you were in a hurry for the first one and you are not for this one. It is optional, you can skip it, and what you type is not sent anywhere: it is written onto the run so whoever opens it later reads what you saw. It shows up as the detail on the timeline's cancelled row.

What stopping actually stops

This is the part worth being precise about.

Nothing interrupts the worker mid-call. Every durable write a run makes is conditional on the run not having settled, so the moment the run is cancelled, the worker's next write loses and is thrown away. Work already inside a call carries on to that next write and then stops mattering.

So the brake bounds what a run can still record, not what a request it has already made will do at the other end. An email that has been sent is sent. This is what the dialog means by "work already sent to another system has been sent", and it is the honest description of what any stop button on any system can promise.

One visible consequence: a stopped run can still gain a line or two of activity from the worker on its way out. The run's status is the truth, and the screen shows the run stopped regardless of what arrives afterwards.

A stop is not a failure

A cancelled run is terminal and neutral. It opens no failure review, it does not count toward the workflow's failure streak, and it does not push a workflow toward demotion. A human pulling the brake says something about that moment, not about whether the process works.

The runs list labels it Stopped by you. The run screen labels it Cancelled. Same thing.

  • Rejecting at an approval checkpoint is not stopping the run. On a codified run a rejection is treated as a correction: the agent is shown what the gate presented and what you said, and asks you again with revised values. See Approvals and gates.
  • Pausing a workflow stops new runs from starting without touching the ones already going. The checkbox above is a shortcut to doing both at once. See Workflows.

On this page