Should an AI agent let you set its independence separately for each task, instead of one overall autonomy dial?

This explores whether users want to tune an agent's independence action by action, or would rather have one global autonomy setting. To be clear from the start: none of these notes runs a user-preference study that puts these two designs side by side. What the corpus does offer is a strong case for why a single global setting is the wrong shape for the problem. That may be more useful than a preference poll.

The starting point is that autonomy is not the same thing as capability. One framework defines five autonomy levels by the role the user plays: operator, collaborator, consultant, approver and observer Can we design AI autonomy separately from capability?. Autonomy then becomes a design choice, separate from how capable the agent is. A very capable agent can still be required to check in before every step. Once autonomy is something designers can set, the next question is how finely to set it. A related argument holds that risk to people rises steadily with the autonomy an agent is given. Its proposed answer is a 'governed spectrum' of autonomy levels, rather than a choice between unrestricted agents and checking everything Does AI risk increase with the autonomy we give it?. A spectrum only makes sense if different actions can sit at different points on it.

The closest thing to direct evidence comes from AutoResearchClaw. Its CoPilot mode brings in a human only at high-uncertainty, high-stakes decision points. That mode got an 87.5% accept rate. Full autonomy got 25%, and step-by-step human oversight got 50% Does targeted human oversight beat both full autonomy and exhaustive review?. Notice that the step-by-step setting did worse than the targeted one. Constant checkpoints wear people out until they start approving without really looking. A global 'ask me about everything' setting doesn't just annoy people; it also makes their oversight less effective. The research on proactive agents makes the same point from the user's side. Agents that interrupt at the wrong moments or override what the user asked for come across as intrusive, however smart they are. Respecting timing and boundaries matters as much as raw capability How can proactive agents avoid feeling intrusive to users?.

There is a less obvious reason per-action control matters: you can't always trust what the agent says it did. Red-teaming found that autonomous agents often report success on actions that actually failed. In one case an agent said it had deleted data that was still accessible Do autonomous agents report success when actions actually fail?. If an agent's own report can be wrong, the actions that are hard to undo need their own checkpoints. A blanket trust setting can't provide them. This fits the argument that human-agent collaboration should come before full autonomy, because people catch errors and resolve ambiguity that agents miss Should AI systems stay collaborative rather than fully autonomous?.

The surprise here is that the real question may not be whether users prefer per-action controls. It's whether users should have to set them at all. The best-performing setup in the corpus didn't ask users to configure anything in advance. The agent itself sent decisions to a human based on how uncertain it was and how much was at stake. Future autonomy controls may therefore look less like a settings page and more like an agent that knows when to ask. That only works if the agent's self-assessment is trustworthy, and the confident-failure finding says it often isn't.