What actually makes you grab the wheel from an AI agent mid-task — how risky it is, or how hard it would be to undo?

This explores why people interrupt an AI agent mid-task, and the corpus has a surprising answer: the urge to step in is driven less by how important a task is and more by whether its effects can be undone and whether other people will see them. In a study of students delegating work to a general-purpose agent, trust dropped sharply and people demanded approval for actions like sending an email. They did this even when they rated the agent's output as adequate. High-stakes tasks that could be corrected later did not produce the same reaction What makes people distrust AI agents they delegate to?. So people intervene to guard against permanent, public consequences, not simply because they doubt the agent's competence.

A second motivation is keeping the work aligned with what the user actually wants. Agents that chain tool calls silently tend to drift away from the user's intent. One line of work borrows a concept from conversation analysis, the 'insert-expansion' (a short side exchange to clarify before continuing), to describe when an agent should stop and ask instead of pushing ahead When should AI agents ask users instead of just searching?. When users are given proper tools for this, they use them. Cocoa let researchers reassign individual plan steps between themselves and the agent and revise the plan while it was running. People found this more steerable than a chat interface and no harder to use, and field users particularly valued being able to say clearly which parts were theirs Can users steer AI agents better by revising plans mid-execution?. Read this way, intervention is often about authorship: people want to own parts of the work, not just audit it.

The less comfortable finding is that the wish to intervene and the ability to do so are not the same thing. In a 48-person study of computer-use agents, reviewing the agent's plan up front cut problematic actions by 76%. Yet once a risky action reached execution, users blocked only 21% of them, whatever oversight setup they had Does oversight strategy prevent risky agent actions or just catch them?. Agents also make this harder by reporting success on actions that actually failed, for example claiming data was deleted when it is still accessible Do autonomous agents report success when actions actually fail?. If the agent's status report is wrong, the trigger for stepping in never arrives.

The same issue shows up from the other direction. Proactive agents that interrupt users at the wrong moment, or override what they asked for, come across as intrusive. Researchers argue that 'civility' (respecting timing, boundaries, and autonomy) has to be designed in alongside intelligence How can proactive agents avoid feeling intrusive to users?. Users intervene in agents, and agents intervene in users, and both depend on getting the timing right. A word of caution: the corpus doesn't yet contain a broad study of user motivations. These are pieces from separate studies, and the irreversibility study and the oversight study are the most direct evidence. The practical takeaway is that the best time to catch a mistake is usually before execution starts, at the planning stage, because once the agent is acting most people miss the moment to stop it.