← Blog
Week of

Retrospective reconstruction · Published · Weekly archive

Actually, use the other one

Long tasks get corrected halfway through. That's normal. Someone remembers a constraint, changes an account, or decides the first option wasn't right after all.

Field NotesRetrospective

Retrospective field note for 2025-08-04. Reconstructed in November 2026 from dated sources, not originally published on the archive date.

Long tasks get corrected halfway through. That’s normal. Someone remembers a constraint, changes an account, or decides the first option wasn’t right after all.

The July ChatGPT agent release makes interruption a product question rather than a chat courtesy. When a person says stop, the system needs to know what has already happened.

Was the message drafted or sent? Was the request prepared or submitted? Was the file changed locally or published? “Stopped” means very different things in each case.

I’d expect the agent to stop starting new actions, check anything already in flight, and give a plain account of the state. If something did go through, say so. Don’t make the person discover it later.

Corrections also need to reach every worker involved. Splitting the task internally is fine. Asking the user to repeat the same change to three agents isn’t. The coordination belongs inside the service.

None of this requires a lengthy apology or a speech about adaptability. Hear the correction, update the plan, preserve the completed work, and continue from the new direction. If the change can’t be undone, that’s the important sentence. Put it first.

Sources

All posts Work with me