Should an AI's eagerness to jump in be something you can turn up or down while you're mid-conversation, not just set once?

This explores which interface controls let people turn an AI's initiative up or down in the middle of a session, not just at setup. The first thing the corpus establishes is that people actually want to do this. In two Slack studies of a brainstorming bot, participants asked it to contribute less often partway through, then pushed it to speak up more once it had gone quiet Do users want to adjust AI proactivity during conversations?. So proactivity isn't a setting you choose once. It changes with the flow of the work. That also overturns an older assumption that initiative is something the system either has or lacks. Plain chat LLMs barely take initiative at all by default Why can't conversational AI agents take the initiative?. The design problem runs in both directions: quieting an agent that talks too much, and drawing out one that defaults to passivity.

The most widely shared control in design practice is the autonomy slider. It's a continuous dial with roughly three stops: suggest only, draft, and execute. Users can set it per task Should agent autonomy be a continuous slider or fixed levels?. The same pattern library goes further and argues that "autonomy" shouldn't be a single dial at all. It breaks it into separate controls: a budget for how much the agent may do, approval gates where a human must sign off, confirm-versus-execute modes, and permissions scoped to individual capabilities Can autonomy controls stay distinct rather than collapse into one toggle?. Both entries are honest about one limit: this is shared vocabulary, not tested evidence. Nobody has yet shown that sliders beat other control shapes, or that people keep adjusting them instead of setting and forgetting.

The strongest empirical answer comes from group settings. Houde's team sorts initiative controls into five axes: *when* the agent speaks, *what* it contributes, *where* (which channel or thread), *who* has authority to change its behavior, and *how* the change is specified What controls help groups manage AI agent participation?. Given a prototype, participants built their own configurations instead of picking presets. That suggests a single slider is too coarse. Someone might want an agent that speaks rarely but substantively, or often but only in a side thread, and one dial can't express either. A separate finding points the same way from a different field. In simulated conversations, monitoring that targeted *which behaviors* to correct cut drift by 87%, while smarter *timing* of corrections added nothing Does monitoring help more by choosing what to correct than when to intervene?. Applied to interface design, that hints that controls over *what* the agent does may matter more than controls over *how often*.

A second family of controls works through the plan rather than through a dial. Cocoa lets researchers assign each plan step to themselves or the agent and revise the plan while it runs. In a lab study this gave more steerability than plain chat without making the tool harder to use Can users steer AI agents better by revising plans mid-execution?. Here initiative becomes a property of individual steps, adjusted right where the work happens. Nielsen's three-layer model explains why that matters. People need an intent layer for saying what they want, an orchestration layer for supervising what the agent is doing, and a direct-manipulation layer for precise fixes How should AI interfaces handle the shift from doing to supervising?. Mid-session adjustment mostly belongs to that middle orchestration layer, which most chat interfaces simply lack.

The underrated piece is architectural. Adjusting initiative mid-session only works if the conversation doesn't freeze while the agent acts. Full-duplex designs like Realtime-Venus keep talking in the foreground while delegated tasks run in the background, all on one shared timeline Can frontends handle delegation while staying conversationally engaged?. That's what makes "hold on, do less of that" possible while work is still underway. In short, the corpus points toward several composable controls, especially per-step assignment and content-focused settings, plus a runtime that stays interruptible. It does not yet have a head-to-head study showing which control wins.