Engineering guide: model agent cancellation as an explicit state machine

On November 1, 2026, the AI section of DEV Community published an engineering guide titled "Agents Keep Running After Stop: Model Cancellation as a State Machine." It was written for developers building Agent products.

The guide describes a failure mode that is easy to miss: the Stop button in a browser can look perfectly honest while the backend keeps working. The chat stream disappears from the screen, but a model request may still be generating, a browser tool may still be clicking, and a background worker may be one step away from sending an email. The author argues this is not a cosmetic bug, but a broken contract. Users read "Stop" as "do not do more work for me," yet most systems only cancel the frontend flow and never translate that request across the other layers involved in a single turn.

To fix this, the author recommends treating cancellation as an explicit state machine rather than a single flag or a signal sent once. The proposed states are running, cancel_requested, cancelling, and cancelled, which make it possible to track where a request actually stopped. Alongside the state model, the guide advises recording what really stopped at each layer and exposing the remaining uncertainty to the user instead of hiding it. In short, a system should be able to tell the user exactly what is still running and what is safe to assume has finished.