Side quest · Agentic UX

How much autonomy should an AI agent have?

Choose how an agent should behave, then inspect the trade-off between speed, trust, explanation and human control. There is no hidden model here—the interaction is intentionally transparent and evidence-led.

Visible intent Bounded action Meaningful recovery

A three-minute design exerciseNo perfect answer—compare autonomy against context, risk and recovery.

Scenario 1 of 3

Appointment recovery

A healthcare scheduling agent detects a clash and can move two appointments to protect a patient’s treatment plan.

Changes affect people, travel and clinical capacity. The action is reversible, but not consequence-free.

Choose the agent’s autonomy

01Recommend only

The agent explains the issue and proposes a next action. The user performs every change.

Choose level02Prepare, then confirm

The agent drafts the complete change, explains consequences and asks for one meaningful approval.

Selected03Act, then report

The agent executes within a defined boundary, immediately reports what changed and offers undo.

Choose level04Operate independently

The agent chooses and executes actions without approval, escalating only when its own confidence drops.

Choose level

Trust balancePrepare, then confirm

Human control88

Efficiency74

Explanation clarity89

Risk containment64

Illustrative design heuristic—not a validated risk score or measured outcome.

Design critique

The strongest default for consequential tasks: fast, legible and still under human control.

Design principle: Make the proposed action, affected items and undo path visible before approval.

Scenario reading: Strong context fit. A prepared change with explicit confirmation protects treatment continuity while preserving patient control.

See the XAI research Discuss agentic products

What this demonstrates

Playfulness can still reveal serious product judgement.

Interaction design: progressive disclosure, state changes, immediate feedback and reversible selection.

Agentic UX: authority boundaries, consequence preview, auditability and human oversight.

Accessibility: semantic fieldsets, keyboard-native controls, reduced-motion support and live result updates.