AI Agent UX
Design experiences for AI that doesn't just suggest - it acts. Booking flights, sending emails, executing code, making purchases. The AUTONOMY framework ensures users stay in control while AI operates at machine speed.
Core Principle
Agentic AI inverts the traditional UX model. In classical UX, users take actions and the system responds. In agentic UX, the system takes actions and users supervise. This requires an entirely new design vocabulary: previews instead of forms, audit trails instead of history, interrupts instead of navigation.
The AUTONOMY Framework
| Letter | Principle | Design Question |
|---|---|---|
| A | Action Preview | Can the user see exactly what the agent will do before it does it? |
| U | User Override | Can the user stop, modify, or redirect the agent at any point during execution? |
| T | Tiered Authority | Does the agent's freedom scale appropriately with risk level? |
| O | Observable State | Can the user see what the agent is doing right now, in real time? |
| N | Narrated Reasoning | Does the agent explain why it chose this action over alternatives? |
| O | Outcome Verification | Can the user verify that the action completed correctly? |
| M | Memory of Actions | Is there a complete, searchable log of everything the agent did? |
| Y | Yield to Humans | Does the agent know when to stop and ask for help? |
The Autonomy Dial
Not all actions require the same level of human oversight. Design a graduated control system:
| Level | Name | Agent Behavior | User Involvement | Example |
|---|---|---|---|---|
| 1 | Suggest | Recommends action, takes none | User must initiate | "I recommend replying with: [draft]. Send?" |
| 2 | Draft & Wait | Prepares the action, pauses for approval | User reviews and approves | Agent drafts email, shows preview, waits for "Send" |
| 3 | Act & Notify | Executes immediately, sends notification | User reviews after the fact | Agent files expense report, sends summary |
| 4 | Act Silently | Executes without notification | User can check audit log | Agent auto-categorizes emails |
| 5 | Full Autonomy | Executes complex multi-step tasks independently | User sets goals, agent reports outcomes | Agent manages a full sales outreach sequence |
Setting the Default Level
| Risk Factor | Default Autonomy Level |
|---|---|
| Action is reversible + low cost | Level 3-4 (Act & Notify or Silent) |
| Action is irreversible OR high cost | Level 1-2 (Suggest or Draft & Wait) |
| Action involves external parties (sends emails, makes payments) | Level 2 maximum until trust is established |
| Action involves sensitive data | Level 1-2 with audit logging |
| User has explicitly granted higher autonomy | Respect the setting, but cap at the risk level |
Design rule: The Autonomy Dial should be a user-facing control, not a hidden system setting. Let users tune their own comfort level.
Pre-Action Design
Before the agent acts, users need to understand what's about to happen.
The Action Preview Card
Every consequential agent action should display a preview card with:
| Element | Content | Required? |
|---|---|---|
| Action summary | One sentence: what the agent will do | Always |
| Targets affected | Who or what is impacted (people, systems, data) | Always for external actions |
| Estimated impact | What changes as a result | For non-trivial actions |
| Alternatives considered | What else the agent could have done | When multiple valid options exist |
| Reversibility indicator | Can this be undone? How? | Always |
| Confidence level | How sure is the agent this is the right action? | When confidence varies |
During-Action Design
While the agent is executing, users need visibility and control.
The Agent Activity Feed
A real-time, scrollable feed showing:
| Component | Purpose | Update Frequency |
|---|---|---|
| Current step indicator | "Step 3 of 7: Sending confirmation email" | Per step |
| Progress visualization | Linear or branching progress bar | Continuous |
| Decision log | "Chose option A because [reason]" | Per decision point |
| Pause/stop controls | Always visible, always responsive | Permanent |
| Skip/redirect | "Skip this step" or "Do X instead" | Per step |
Critical rule: The stop button must work instantly. If the agent can't be stopped mid-action, the UI must make this clear BEFORE the action begins.
Post-Action Design
The Action Receipt
After every consequential action, provide an action receipt:
| Element | Content |
|---|---|
| What was done | Plain-language summary of completed actions |
| What changed | Before/after diff of affected data |
| Undo option | One-click reversal if action is reversible |
| Time-limited undo | For partially reversible actions: "You have 30 seconds to undo the email send" |
| Audit link | "View full activity log" |
The Undo Architecture
| Action Type | Undo Pattern | Time Window |
|---|---|---|
| Data modification | Full revert to previous state | Unlimited |
| Message sent | Recall + replacement option | 30 seconds to 5 minutes |
| File deletion | Soft delete → permanent after 30 days | 30 days |
| Purchase/payment | Cancellation request + refund flow | Varies by provider |
| Multi-step sequence | Step-by-step rollback with checkpoint selection | Per step |
| Irreversible action | Prevention-only (no undo, must block before execution) | N/A |
Trust Ramp: Progressive Autonomy
New agents should not start at Level 5. Build trust incrementally.
The Trust Ramp Pattern
| Phase | Duration | Autonomy Level | Unlock Condition |
|---|---|---|---|
| Onboarding | First 5 interactions | Level 1 (Suggest only) | User completes first review cycle |
| Building | Next 20 interactions | Level 2 (Draft & Wait) | >80% approval rate on previews |
| Established | Ongoing | Level 3 (Act & Notify) | User explicitly opts in |
| Trusted | After sustained use | Level 4 (Act Silently) | User configures specific action types |
| Delegated | Expert users only | Level 5 (Full Autonomy) | User sets goals + boundary conditions |
The cardinal rule: Users can always dial DOWN autonomy instantly. Dialing UP requires confirmation.
Multi-Agent Coordination UX
When multiple agents work together:
| Challenge | Design Solution |
|---|---|
| Which agent did what? | Color-coded agent identity in the activity feed |
| Agents disagree on approach | Surface the disagreement to the user with each agent's reasoning |
| Agent A's output feeds Agent B | Show the handoff explicitly: "Research Agent found 5 candidates → Outreach Agent will draft emails for your review" |
| Cascading failures | If Agent A fails, halt Agent B and notify user before damage propagates |
Anti-Patterns
| Pattern | Why It Fails |
|---|---|
| "The AI will handle everything" messaging | Sets unrealistic expectations; users feel betrayed at first failure |
| Burying the stop button | Users feel trapped. Panic → permanent distrust |
| Asking permission for every micro-action | Decision fatigue. Users disable the agent entirely |
| No audit trail | "What did it do while I was asleep?" Users can't trust what they can't verify |
| Irreversible actions without explicit consent | One wrong autonomous action and the user never delegates again |
| Showing agent "thinking" animation without substance | A spinner is not transparency. Show what the agent is actually evaluating |
Quick Reference
| Task | Framework Element | Key Deliverable |
|---|---|---|
| Design a new AI agent product | Full AUTONOMY framework | Autonomy Dial settings + Pre/During/Post action flows |
| Add agent capabilities to existing product | Trust Ramp Pattern | Phased rollout plan with unlock conditions |
| Audit agent UX for safety | Undo Architecture + Action Receipt | Gap analysis: which actions lack undo, preview, or audit? |
| Design multi-agent coordination | Multi-Agent Coordination table | Agent identity system + handoff visualization |
| Set autonomy defaults | Setting the Default Level table | Risk-matched autonomy levels per action type |
Integration
Works with: ai-trust-transparency (explaining agent decisions), ai-error-resilience (agent failure recovery), ai-safety-guardrails (agent action boundaries), ai-feedback-loops (rating agent performance), ai-ux-eval (scores the predictability dimension).
