A research team built a working test environment called Control OSWorld to measure how well humans catch harmful agent actions on a desktop computer. The setup gave human monitors a live view of an agent working through tasks, then measured how often the monitor caught a dangerous move before it landed. Monitors missed the harmful action 63% of the time. The result puts a hard number on something designers have assumed but seldom tested: watching an agent work does not produce safe outcomes on its own. The study also varied how much information the monitor received. Monitors who saw only a screenshot caught fewer problems than those who also saw the agent's planned next step. That gap matters for product teams deciding what to show a human when an agent is mid-task. More information raised catch rates, but the ceiling stayed low, and response time became a bottleneck. Agents often moved faster than a monitor could form a judgment and act. For teams building agent products, the finding reframes the oversight question. A live view is not the same as meaningful control. A human who watches an agent run may not have enough time or decision support to redirect it. The paper argues that the design of the oversight interface, not the human's attention, is the variable worth optimizing. That shifts responsibility toward product teams who decide what information the monitor sees and when.