Context card
Pause, abort and undo: three functions in agent execution
What is the difference between pausing, aborting and undoing an action an AI agent has started?
The short answer
Pausing, aborting and undoing are three separate functions of the execution environment around an AI agent. Pausing and aborting only stop further steps; neither takes back an effect that has already left the system, such as a submitted form or a sent message. Undoing is a separate counter-action in the target system and exists only if that system offers one.
For: Technical writers and engineers documenting AI agent systems
Key points
- Pausing keeps context, sessions and permissions alive; the run can continue.
- Aborting should also withdraw the working credentials and release the tools.
- Undoing requires a cancellation path in the target system and its own authorization.
- After a timeout the execution layer knows what it sent, not what the target did.
- Resending is a second action, not a repetition of the first.
The context
Stopping a run is not reversing an effect
An AI agent runs inside an execution environment that splits a task into steps: it plans, calls a tool, reads the response and decides the next step. A stop command reaches the loop that issues the next step. It does not reach a request already on the wire, a job another service has accepted, or a record a target system has already written. Whether an effect can still be withdrawn is a property of the target system, not of the agent, and it has to be designed in like any other control function.
Pause, abort, undo
Pausing holds the loop and keeps context, open sessions and issued permissions. The run is resumable, but those permissions stay usable while it is parked. Aborting ends the run and should also withdraw the working credentials and release the tools; a call whose result never arrived must be marked as unknown rather than repeated. Undoing is a new action in the target system — a cancellation, a correction, a deletion request. It needs its own authorization and only exists where the target offers it; many public submission forms offer nothing comparable.
Unknown results and duplicates
After a timeout, the execution layer knows the request it sent, not the state it caused. Sending again is a second action. Methods such as PUT and DELETE are defined as idempotent in HTTP, a form POST is not; where the target service accepts a client-supplied identifier, a second attempt can be recognized as the same transaction. Without such an identifier, the defensible default is to stop and hand the case to a person.
What the documentation has to state
For every consequential action type, the documentation should name which of the three functions exists, who may trigger it, how long it takes effect and what stays in place afterwards: submitted data, created records, notified third parties. Run records that show the request, the response and the human intervention need access protection, a defined retention period and data minimization — a security requirement of its own, not a by-product of logging.
Questions readers ask next
- Does stopping the text generation stop the action?
- No. Generation and execution are separate. A tool call that has already been dispatched continues, and a background job started earlier keeps running until its own controller stops it.
- Who should be able to abort a run?
- A named role with an access path that does not depend on the agent's own tooling. If the only stop button is inside the agent's interface, a stuck or misdirected run can be hard to reach.
- Does a confirmation dialog cover this?
- A dialog authorizes an action before it happens. It says nothing about stopping a run in progress or reversing a completed one; those are different functions and need their own description.
Sources
- RFC 9110: HTTP Semantics — RFC Editor (IETF), June 1, 2022
- Agentic AI – Threats and Mitigations — OWASP GenAI Security Project, October 13, 2026
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.