Last reviewed
Correct answer: B. Because some actions are hard to reverse, like deleting data or sending a message
Explanation
Reversibility is the axis. Reading a file, running a search or listing a directory can be got wrong cheaply; the cost is a wasted turn. Deleting a table, moving money, or sending a message to a real person cannot be undone by trying again. Google warns against computer use for actions where serious errors cannot be corrected, and OpenAI's guidance for computer-use agents lands in the same place: let the agent do as much safe work as it can, then pause exactly when the next action would create external risk.
Production permission systems draw that line in code. A Gemini computer-use response can carry a safety decision marking an action require_confirmation before your client runs it. OpenAI names deleting data, changing sharing settings or API keys, and sending anything to a third party as always-confirm.
Notice how narrow the real rule is. Confirmation targets consequential actions; it is not requested per generated word, which would make the system unusable, and it is not demanded for every tool call. An agent that pauses is obviously capable of acting. Pausing is a judgement about when, not an inability.
Sources
“The safest default is to let the agent do as much safe work as it can, then pause exactly when the next action would create external risk.”
“We recommend supervising closely for important tasks, and that you avoid using the Computer Use capability for tasks involving critical decisions, sensitive data, or actions where serious errors cannot be corrected.”
“Agents can then pause for human feedback at checkpoints or when encountering blockers.”
Practise 3 questions on this topic
Take AI Agents Basics — Timed Test (3 questions) — scored instantly, explanation for every question, no login.