Traditional AI UX is conversational: the user asks, the AI answers, the user decides what to do. Agentic UX is collaborative: the AI proposes actions, the user approves or redirects, the AI executes. The interface patterns that work for chatbots don't work for agents.
The trust lifecycle
Trust in AI follows a predictable arc. Designing for each stage prevents the common pattern where users start enthusiastic and disengage after the first bad experience.
Formation (first 2-3 interactions). Users build their mental model of the AI in the first few encounters, and that model is sticky. Design implication: the first interaction must deliver real value, and the AI should show a boundary early. An AI that admits what it can't do earns more trust than one that tries everything.
Calibration (first week). Users test the AI against their expectations. Consistency matters more than occasional brilliance here. A reliably good AI earns trust faster than an unpredictable excellent one. This is when users discover failure modes - your error recovery design determines whether they bounce or adapt.
Reliance (ongoing use). Users develop habits. Over-trust and under-trust both emerge. Monitor for users who stop checking AI output (they need gentle friction) and users who manually redo AI work (they need evidence of reliability).
Recovery (after failure). Every AI will fail. Trust recovery requires more design effort than trust formation. Acknowledge failures explicitly - silent corrections erode trust. Recovery takes multiple positive interactions, not one apology.
Design principles for agentic interfaces
Make the AI's confidence visible
Users build wrong mental models: "the AI knows what I mean," "the AI is always consistent," "the AI is either right or wrong." Design against these misconceptions.
When the AI is uncertain, show it. When it's working with incomplete information, say so. Don't let ambiguous inputs produce confident outputs. Calibrated confidence is a feature, not a weakness.
Design for the reversible case first
Every agent action should be reversible when possible. For irreversible actions (sending messages, deleting records, publishing content), require explicit confirmation with a preview of exactly what will happen.
The pattern: propose → preview → confirm → execute → undo window. Skipping any step increases the blast radius of errors.
Progressive disclosure of capability
Don't show users everything the agent can do on day one. Reveal capabilities as users demonstrate comfort with simpler actions. This maps directly to the progressive trust model in agent experience: read-only → draft → send with approval → autonomous.
Each capability reveal should come with context: what it does, what it doesn't do, and how to undo it.
Error recovery over error prevention
You cannot prevent all AI errors. Design great recovery instead:
- Make errors visible immediately, not buried in logs
- Show what the AI did and why it thought that was correct
- Provide a one-click path to the correct state
- Learn from the error (if your system supports feedback loops)
The worst agentic UX pattern: the AI takes an action, the action is wrong, and the user has to figure out what happened and how to fix it manually.
Common anti-patterns
The chatbot shell. Wrapping agent capabilities in a chat interface because that's what the team knows how to build. Chat is good for exploration and Q&A. It's poor for action confirmation, multi-step workflows, and status monitoring. Use the right interaction pattern for the task.
The confidence theater. AI that always sounds certain, even when it's guessing. Users calibrate trust based on how the AI presents information. False confidence creates over-trust that shatters on the first wrong action.
The permission dump. Asking for all integrations and permissions upfront. Users grant broad access, forget what they gave, and then blame the AI when it acts in unexpected contexts. Request permissions incrementally, in context, with clear explanation of why each one is needed.
The invisible action. AI that takes actions without surfacing what it did. Even autonomous agents need an activity log that users can review. The goal is informed trust, not blind trust.
Connecting the dots
Agentic UX sits at the intersection of agent experience (how the agent is designed), security thinking (what the agent is allowed to do), and user-centered design (what the user actually needs). Getting any one of these wrong undermines the others.
The AI Health Indicator assesses Tone and Alignment - two dimensions that are directly shaped by UX decisions. A technically excellent agent with poor UX will score low on both.
Related practices
Want help with agentic ux?
I coach teams on this practice. Let's talk about your situation.
Get in touch