I judge a long agent run by what I asked for against what came back, and by how many times I had to step in. Hours of runtime don't count on their own. On July 9, 2026, OpenAI launched ChatGPT Work for long tasks across apps and files. The promise of continuity appeals to me because I want to leave one task moving while I handle another.
That raises what I expect from the return. A long log of actions helps when I need to investigate a problem. When I reopen the work, I first want to know where it stands: is the request done, or is a specific doubt blocking it? If I have to read everything from the start to find out, delegating just gave me more work.
The original task also has to survive the hours. An investigation can find an interesting direction and drift away from the request while it digs in. I want to understand how that helps the delivery. A side idea can be written down for later; it doesn't get to silently replace what I wanted finished.
I also need to know early when the run depends on me. Some decisions will need my answer. If they're predictable, spotting them during preparation lets me plan the work around them. An agent that runs for hours but calls me the whole way needs me available in a way I didn't expect when I handed it the task.
So when I come back, the return needs four things: the current state, done or blocked by what; the decisions waiting on me, written as questions; any drift from the original request, recorded; and evidence of what moved forward.