The two modes can coexist in one product, but switching should be the user's decision, not the system's

What it is

Tool and agent are not an either-or product shape: one editor can offer step-by-step retouching tools alongside a one-tap "enhance" delegation. The question is not whether they coexist but who owns the switch—when users choose explicitly, the modes complement each other; when the system takes over silently, users lose track of who is in control.

Why it happens

Control expectations shift with context: under time pressure or low risk users delegate; when learning, facing high stakes, or wanting to express themselves, they hold the steps. These factors are only partially inferable—and the evidence is often a proxy signal. A wrong inference means the agent seizes control and the user can no longer find their own path, or a suddenly intercepted tool interrupts skill execution. The sense of mode ownership depends on switching being predictable.

Studying it

Compare switching strategies: explicit user switching, system-initiated switching based on task features, and system suggestions with user confirmation. Measure mode misattribution (the agent steps in while users believe they are in control, or vice versa), correction time, and trust and control scales. In logs, clusters of "undo immediately after a switch" expose where the system misjudged.

Where it stops holding

Explicit switching does not forbid sensible defaults: agent-first with an easy path to tools is reasonable for brand-new users, and remembering last choice beats re-asking for frequent ones. In safety-critical or irreversible flows, automatic system intervention requires separate confirmation and cannot ride on a default. Switching itself has a cost—asking every time is its own burden.

Applying it

• Give every dual-mode flow an explicit entry ("I'll do it / hand it to the assistant") and keep the current mode continuously visible.
• When suggesting a switch, state the reason and scope and wait for confirmation; never change control mode silently.
• Remember per-task-type mode preferences and minimize repeated prompts.
• Acceptance: verify every switch leaves an undo path, and drive system-initiated switches to zero.

Related

• Same group: B4.16.1 In instrumental interaction the user controls each step; in agentive interaction the user delegates a goal · B4.16.2 The two modes demand different feedback: tools need immediate response, agents need visible process · B4.16.3 Delegation transfers control and responsibility together; its boundary must be stated before the handoff · B4.16.5 Agentifying operations users could do directly removes space for skill expression
• Nearby: B4.17 Interaction frameworks and state transition models · L3 Intelligent System Interaction
• Search terms: mixed-initiative interaction · mode switching · user control