Should an AI agent have to pause and ask before it acts, or just act and tell you afterward?

This explores where a user's control over how much an agent does on its own belongs in Nielsen's three-layer model. The short answer is that Nielsen never names autonomy as a control. But his middle layer is where it effectively lives. How should AI interfaces handle the shift from doing to supervising? describes an Intent Surface, where you state the outcome you want. Next comes an Orchestration Surface, where the agent shows its plan, asks for consent, and afterwards gives you a 'receipt' of what it changed and what can still be undone. Last is a Direct-Manipulation Surface: the ordinary point-and-click interface, kept as a fallback for fine corrections and emergency overrides. Autonomy is really a question of how often the agent has to stop and ask before acting, and in Nielsen's model that asking happens at the orchestration layer.

The design community's version of this control makes the placement clearer. Should agent autonomy be a continuous slider or fixed levels? describes a per-task 'autonomy slider' with three stops: suggest only, draft, and execute. Each stop is a different setting for the consent step. At 'suggest' the agent proposes and you decide. At 'draft' it prepares the work and you approve it. At 'execute' it acts and reports back afterwards. Seen this way, the slider sets how heavy or light the Orchestration Surface is for a given task. Can we design AI autonomy separately from capability? fills in the human side. It defines five levels by the role the user plays: operator, collaborator, consultant, approver, or observer. It also makes the most useful point here: autonomy is a design choice that is separate from how capable the agent is. A very capable agent can still be required to check in before every step. So autonomy isn't a property of the agent that the interface has to put up with. It's a setting the interface chooses.

One part of autonomy arguably belongs earlier, at the Intent Surface. Should agentic AI interfaces require users to specify risk tolerance upfront? argues that before an agent runs, users should spell out their risk tolerance, the actions it must never take, and the conditions it needs permission for. That turns the boundaries of autonomy into part of stating what you want. Read alongside Nielsen, the boundaries are set at the intent layer, the moment-to-moment checkpoints run at the orchestration layer, and the direct-manipulation layer is where you take control back when the first two fail.

Why does this placement matter? Does AI risk increase with the autonomy we give it? argues that harm grows steadily as agents get more autonomy, and that a controlled range of levels is safer than either unrestricted agents or checking everything. Nielsen's layers offer one way to build that controlled range into an interface. There's a quieter risk too. Does attributing agency to AI actually reduce human sense of control? finds that people feel less in control once they credit an AI with judgment of its own, even when they still hold formal authority. A slider set to 'execute' could speed up that feeling. Nielsen's post-action receipts and his aim of calibrated trust can be read as a counterweight: they keep you aware of what was done on your behalf.

A caveat: Nielsen's model is a practitioner's design proposal, not a tested framework. Nobody in this collection has studied where autonomy controls work best across the three layers. The unexpected takeaway is that autonomy may not need its own layer at all. It can be treated as a setting for how much negotiation the orchestration layer requires, with its limits fixed earlier at the intent layer.