What changes when an AI workflow takes minutes instead of milliseconds, and the simple pattern that keeps it reliable.
A thought to take with youPut long-running work on a queue, show progress to the person waiting, and make every step safe to retry.
Why requests break
A web request is built to finish in seconds. A multi-step AI task can take minutes: it reads documents, calls tools and waits for other systems. Somewhere along the way a timeout or a dropped connection cuts it off, and nobody is sure how far it got.
The pattern we use
The request only starts the job and returns straight away. The work runs on a background queue that saves its progress after each step. The person sees live progress in the interface, and can leave and come back.
Each step is written so it can run twice without doing harm, because sooner or later it will.