AI agents are taking longer jobs. The return screen matters more than the chat.

OpenAI’s August enterprise report says organizations are asking AI to carry out more substantive delegated work, not just answer questions. Its June research describes many sampled users making requests estimated to exceed an hour of human work, while cautioning that those time estimates are directional. Either way, the ordinary experience is changing: people are no longer only reading an answer. They are returning to work that has been moving without them.

A chat window handles the first ask well. It is much worse at the return. The person who opens it later is usually looking for five plain things: what the job was meant to do, what it finished, what it changed, where it got stuck, and whether there is a decision waiting. If those answers are buried in fifty tool calls, the assistant has created a new kind of catching-up work.

A recent AgentGUI preprint makes this distinction concrete. The project groups activity, task details, files and debug material so a person can watch or redirect a long-running agent without reading every raw event. In its eight-person comparison study, participants answered questions about agent trajectories faster with that interface than with the comparison dashboard. That is a small research sample, not a universal design verdict. It is still a useful reminder that visibility is not the same thing as legibility.

The better return screen is probably smaller than most agent dashboards. Start with a one-paragraph result: “Compared eight vendor policies; drafted the three-page brief; two sources disagree on the deadline; I did not send anything.” Put the changed files or links beside it. Then give the person an obvious choice: review the draft, answer the open question, keep going, or stop.

The full history should remain available. When an output looks wrong, a trace can show the source, the tool action or the assumption that caused it. But making every user begin in the debug view is like leaving a mechanic’s work order on the steering wheel instead of telling the driver whether the car is ready.

This is less about making an AI assistant look busy. It is about lowering the cost of interruption. Someone gets pulled into a meeting, closes a laptop, covers a coworker’s task or simply loses the thread. The product should not punish that ordinary day by making them reconstruct an invisible sequence of actions.

A practical test: hand the returned task to someone who did not start it. Can they say what is done, what is uncertain and what happens if they do nothing? If not, the status is still written for the system, not the people who have to live with its work.

AI agents will keep taking longer jobs. The calmer product move is not another stream of activity. It is a clean place to come back to.