An agent busy for hours interests me when I return and find that work has advanced. On July 9, 2026, OpenAI introduced ChatGPT Work for long tasks across applications and files. That promise of continuity appeals to me because I want to leave one activity progressing while I deal with another.
In June 2026, I had already requested persistent agents and work distribution aimed at speed. On July 12, I questioned whether I was using the available capacity across Cursor, Codex, and Claude Code. I had tools at my disposal and wanted to know whether they were being used well. Idle capacity bothers me too, particularly while tasks are still waiting.
That increases my expectations for the return. A long history of actions can help investigate a problem. When reopening the work, I first want to understand its current state: is the request finished, or does a concrete uncertainty prevent completion? If I have to read everything from the beginning to find out, delegation has handed me a new obligation.
In a hypothetical situation, execution discovers that more preparation is necessary. I expect it to advance independent work and formulate the decision it cannot make. Then I return to the outstanding question alongside what has already been resolved. Spending the entire period awaiting my answer while useful work was possible wastes the continuity that made me delegate.
The original task also needs to survive the duration. An investigation can discover an interesting direction and drift away from the request while exploring it. I want to understand how that helps delivery. An additional idea can be recorded for later; it does not acquire permission to silently replace what I wanted finished.
I also need early notice when execution depends on my presence. Some decisions can require my answer. Identifying predictable ones during preparation lets the work be organized around them. An agent that lasts hours but calls me throughout the process demands a different availability from what I expected when assigning it the activity.
I will judge long work by comparing the request with the result and counting the interventions required along the way. Active runtime alone cannot settle that calculation. I want to return to an understandable situation, with evidence of progress and the right question formulated when a decision from me is missing. If available capacity helps get there, good. If it only keeps a window occupied longer, execution still needs organizing.