Should agent autonomy be a continuous slider or fixed levels?

The AI UX Playground glossary defines an autonomy slider as a control that lets "users set how independently an agent may act, from suggest-only to draft to execute, often per task or surface." The entry names a specific UI shape, a slider, for the degree of agent initiative, and specifies that the setting can vary by task and by surface rather than being fixed once for an entire product.

The three named points, suggest-only, draft, execute, describe graduated levels of agent initiative, where each step up removes one layer of human confirmation before an output takes effect: a suggestion requires the user to act on it, a draft requires the user to review and send it, and execution requires neither. By specifying the control as "per task or surface," the glossary implies that the right level of agent independence is not a fixed property of the user or of the agent, but of the specific job and context — a person might want suggest-only for a billing flow and execute for calendar scheduling, in the same product.

This sits in contrast to Does machine agency exist on a spectrum rather than binary?, whose five levels (passive, semi-active, reactive, proactive, cooperative) describe a taxonomy of machine types rather than a control a user adjusts directly. The autonomy slider turns that descriptive spectrum into a prescriptive, adjustable interface element, collapsing it to three settings a person can choose per task. It also bears on How should AI interfaces handle the shift from doing to supervising?: a control of this kind would live in Nielsen's orchestration surface, where users negotiate and consent to agent plans, giving that layer one concrete mechanism for setting how much consent an agent needs before acting. It also changes what an Should agent accountability be designed for end users instead of engineers? needs to carry: at the execute setting, the receipt becomes the only record a user has of an action that no confirmation step preceded.

The glossary does not say how finely grained the slider should be, whether three settings suffice across most agent tasks, or how a product should handle a mismatch between a user's chosen autonomy level and an agent's actual confidence or competence on a given task. It names a pattern rather than evidence; it does not show that a slider outperforms other control shapes, such as toggles or per-action approval queues, for setting agent independence. The implication the excerpt supports is only that the design community has converged on a vocabulary for continuous, per-task autonomy control, not that this particular control shape has been validated.