Should agent control be one setting or several separate controls?

aiverse's pattern library, a catalog of named interaction patterns for AI products, lists three entries that each give a user a different way to limit an agent's latitude. "Guardrails & permissions" is defined as "Boundaries, controls, and permissions that show what AI can and cannot do." "Autonomy control" is "Granular controls over how independently an agent can act on the user's behalf." "Pause & Stop agents" is "Give users immediate control to pause or stop an agent mid-execution." The library keeps these as three distinct catalog entries rather than collapsing them into one agent control pattern.

The three also sit at different points in the library's running order rather than next to each other: guardrails & permissions appears near patterns about onboarding and framing what the AI can do (branding, mode selection, disclaimer); autonomy control appears near input-stage patterns (structured input, prompting assistance, temporary chat); pause & stop agents appears later, grouped with review and correction patterns (review actions, error states & recovery, evaluation/testing). The ordering implies each pattern answers a question tied to a different moment in the interaction — what the agent is permitted to do before it starts, how independently it acts once started, and whether it can be interrupted mid-task — rather than one configuration screen covering all three.

This maps onto the four-part oversight framework in Does human-in-the-loop design actually enable effective oversight?: guardrails & permissions and autonomy control correspond to that primer's "control" component (meaningfully directing what the agent can do), and pause & stop agents corresponds to intervention (pausing, redirecting, or stopping it in time). That primer's fieldwork found builders supply intervention and some control while downplaying the knowledge and observation components (named separately in Should agent trust signals stay separate or merge together?); aiverse's library shows that practitioner design vocabulary already distinguishes three separate control- and intervention-side patterns, consistent with that being the half builders prioritize. It also bears on How can users verify what an AI agent actually plans to do?, which argues a user needs a way to check an agent's plan before trusting any steering control — autonomy control and pause & stop are the steering affordances that verification would need to inform, but the library's definitions do not say what, if anything, feeds them.

The library gives one-sentence definitions with no implementation detail, no evidence on how often products ship all three patterns together, and no measure of whether three separate controls help users or just add configuration surface. Its "what AI can and cannot do" framing for guardrails & permissions is also a boundary on action, not a check on whether an allowed action is safe — a gap Do overseers focus on agent execution or action safety? finds in practice even where a plan-correctness check exists. What the excerpt supports is a naming claim: the design community has separate, named vocabulary for bounding, throttling, and interrupting agent autonomy. It does not show that products implement all three, or that users benefit from having them split apart rather than unified.