Why does every AI agent framework eventually reinvent a workflow engine?

I keep coming back to the point where an agent stops being a loop and starts being a process that has to survive without the original Python process. Suppose it has retrieved some documents, called two APIs, and is waiting for approval before making a change. If the worker disappears, "run the prompt again" is a pretty strange recovery strategy. SO the next model response may choose a different path, and the APIs may already have done something. At that point I want a record of completed work, durable waiting, a retry policy, and a way to resume. That's a very familiar set of workflow-engine requirements. "Every" is doing some work in the title. A short, read-only agent may not need much of this. LangGraph already has persistence and interruption mechanisms; Temporal starts from durable workflows; a custom queue-and-database setup can implement a narrower contract. These aren't interchangeable, but they're converging on some of the same questions. Diagrid Catalyst interests me as another way of drawing that boundary:KEEP THE AGETN FRAME WORK, put durable execution underneath it. I wouldn't take that to mean an external side effect becomes safe to repeat automatically. The application still has to define what counts as the same operation. My preference would be to let the model choose useful next steps while keeping the execution record and recovery rules explicit. The part I'm unsure about is ownership: should the agent framework own that durable lifecycle, or should it integrate with an execution system that can also run ordinary non-AI jobs???