Unattended is not a different kind of process; it is the same workflow with the human removed from the driving seat. That framing matters because it determines what you must build: not a second system, but a dispatch path (who runs the agent), a selection mechanism (what it works on), and reliability plumbing (what happens when the backend is down).
The reliability part is where naive implementations fail quietly. Dispatch must survive crashes between systems (written transactionally, redelivered idempotently), and a run that dies must be reaped rather than leak.
Judged well, unattended execution is measured in receipts: each run should record its session, its retries, and its cost, so autonomy shows up as a line item you can read rather than a bill you discover.
There is also the question of what a run does when it needs a person. A run with nobody watching cannot ask in the chat. The robust answer is that the process itself contains the stopping points: the run reaches a gate, that piece of work waits there, and a person answers when they can, while other work keeps moving. Without that, the run either guesses or stops for good.