AgentApprovals
← All guides

You stopped the agent. Did its action stop?

Separate cancellation requests from stopped execution and verified effects. Use a worked batch example and an editable outcome worksheet before retrying.

Your agent is updating three delivery records. You notice the wrong date and press Stop. The spinner disappears. Can you assume the records stayed unchanged?

No. A cancellation request tells the system what you want to stop. You still need evidence of what stopped and what already happened.

This guide provides a proposed procedure for that investigation, with a fictional batch example. It does not test a particular agent host. Download the cancellation outcome worksheet to keep the request, execution status and destination evidence separate.

Keep three observations separate

ObservationWhat it establishesWhat it leaves open
The interface stopped waitingThe local interaction endedWhether the remote operation is still running
A service acknowledged a cancellation requestThe service accepted the request according to its contractWhether execution stopped before any effect
The destination shows a changed recordThat inspected record has the observed valueWhether other records changed, or work remains queued

These are evidence categories, not names that every product uses. Read the actual host and service contract before interpreting a button or status.

For example, the MCP cancellation specification dated 18 June 2025 allows a receiver to ignore a cancellation notification when processing has finished or the request cannot be cancelled. It also describes races between cancellation and completion. That named revision does not establish what your client implements.

The separate MCP Tasks extension dated 28 July 2026 defines cooperative cancellation through tasks/cancel. Its acknowledgement is not a promise of a cancelled terminal status. Inspect the task’s eventual state and result. Neither protocol contract promises to reverse an earlier business effect.

Follow one interrupted batch

Imagine this sequence. It is a design example, not an observed vendor incident.

  1. Record A changesThe destination saves Wednesday.
  2. The owner requests StopThe client asks the service to cancel the batch.
  3. The worker stopsRecords B and C remain unchanged in this example.
Stopping the remaining work does not restore record A.

Suppose the service now reports the worker stopped, and a permitted inspection of all three records confirms that outcome. The accurate report is: “The worker stopped after changing A. B and C were unchanged at inspection.” Saying only “cancelled” hides the partial result.

If you have only the local Stop indication, the outcome is still unknown. Do not invent the rest of the sequence. An unchanged record seen while the worker is still active also does not prove it will remain unchanged.

Decide from evidence, not the spinner

Preserve the operation’s identity before investigating. Connect the approval or proposal ID, tool request, remote job and destination records where the integration provides those identifiers. Keep them in an authorized private record; some task handles can grant access and should not appear in a public incident post.

Then use this proposed decision table:

Evidence availableReportNext step
Only local waiting endedRemote outcome unknownQuery the authorized service status or ask its operator
Cancellation requested; worker still activeStop requested, execution unresolvedFollow documented status checks; avoid a duplicate operation
Worker is stopped; some writes are confirmedStopped with partial effectsIdentify affected records and prepare a separately authorized repair
Terminal execution and complete destination evidence establish no effectNo effect observed within the verified scopeReassess the desired action and its authorization before any new attempt
Worker completed despite the requestCompleted after cancellation was requestedInspect the result and decide whether a correction is needed

Missing, inaccessible or expired status data belongs in the unknown category. Do not treat the absence of a receipt as proof that nothing happened. For a large batch, a sample of records cannot establish the outcome of every item.

A repair is another action. Restoring an old value could overwrite a legitimate later edit. In the example, first inspect A’s current revision and show the proposed correction to the authorized reviewer. The lifecycle guide’s conditional-write example shows how to avoid overwriting a newer version.

Give an agent a bounded investigation

The worksheet asks for observations and their sources before a proposed next action. An agent can summarize supplied logs, match identifiers and point out gaps. Those judgments do not create permission to read another account, cancel another job or issue a compensating write.

Deterministic checks can compare IDs, timestamps and known record versions. They cannot infer a successful cancellation from a missing response. An authorized operator decides the scope of investigation and any repair. Keep the original logical operation identity when reconciling; creating a new request ID does not erase a possible earlier effect.

For a test environment, use synthetic records and separately authorize the exercise. Observe cancellation before work starts, after one write, and after completion. Record both execution status and the destination for each case. We have not run those integration tests here. The useful result is an honest outcome record, including “unknown” when evidence cannot resolve it.