Thought experiment: you hand a task to someone who can never say "that part's done"

Say you give a coworker a list of forty small jobs and leave for the night. They can work as long as they like, but they have one odd limitation. Nothing they finish ever gets crossed off, in their head or on paper. Every job stays on the table next to the one they're doing. By job twenty-five, a lot of what's on the table is finished work. They'll look at it and wonder whether it still needs attention. Some of it they'll redo, probably well, probably slightly differently than the first time. That's roughly what an unattended agent does on a long multi-step run. Picture a refactor where it renames an interface and updates every call site. A few hours in, it may re-update files it already got right, because nothing marked them closed. Every tool call and every finished step is still in context, weighted about the same as the step it's on now. What I find interesting is why this shows up less when a person is in the loop. In a chat you move on after each step, and your next message quietly tells the model that thing is settled. Nobody assigned you that job, you just do it. Take the person out and nothing picks it up. I don't think a bigger context window fixes it, since that's more room for finished work to sit in. Something explicit seems closer, like a task list the agent has to update before moving on, or collapsing a finished step into a one-line note. I haven't tested which holds up better. If anyone here runs overnight jobs, I'd like to know what you ended up using.