Designers love a smooth 'autonomy slider' for how independent an AI agent is — but has anyone actually proven it beats a simple on/off switch?

This explores whether there's actual evidence that a continuous 'autonomy slider' gives users better control over how independently an AI agent acts than a simple on/off toggle. The short answer: the corpus doesn't contain a head-to-head test. What it does show is that the design community has settled on the slider, and that the slider is less continuous than its name suggests. The AI UX Playground glossary names the autonomy slider as a per-task control, but in practice it collapses into three stops: suggest-only, draft, and execute Should agent autonomy be a continuous slider or fixed levels?. So the real comparison isn't slider versus toggle. It's two settings versus three, and the source itself notes that this pattern hasn't been checked against any other control shape.

The same pattern library pushes past the single-dial idea. It breaks autonomy into several separate controls: spending or action budgets, human approval gates, suggest/confirm/execute modes, and permissions scoped to each capability Can autonomy controls stay distinct rather than collapse into one toggle?. That framing suggests both the slider and the toggle may be the wrong shape. 'How independent should the agent be?' is really several questions: independent at what, with how much money, and with which tools. Here too the library offers shared vocabulary, not evidence. It doesn't show that any of these controls stops people from rubber-stamping approvals or losing oversight once a product ships.

The closest thing to empirical support comes from a different angle. Cocoa let users assign each step of a plan to themselves or to the agent, and revise those assignments mid-task. In a 16-person lab study, users found it more steerable than a plain chat interface and no harder to use Can users steer AI agents better by revising plans mid-execution?. That's evidence that fine-grained, step-by-step delegation beats having no explicit control at all. It isn't evidence about sliders, and it hints that the useful unit of control may be the individual task step rather than a global setting.

The governance literature explains why the design question matters. One line of argument holds that risk to people rises steadily with the autonomy handed to an agent, and recommends a governed range of autonomy levels over either full independence or approving every action Does AI risk increase with the autonomy we give it?. That supports having graded levels in principle. A heuristic review of nine deployed agents found that principles like keeping the agent's intent visible and preventing dependency were most often violated in companion and advisory products Do productivity agents actually protect users better than emotional ones?. So whatever control shape you pick, its value depends on whether users can see what the agent is about to do.

To sum up: the claim that sliders outperform toggles is currently a design convention, not a research finding, at least in this collection. The more interesting open question the corpus points toward is whether autonomy should be one dial at all, or a set of separate controls tied to specific steps, budgets, and capabilities.