10 Heuristics for
Agentic Experience
Inspired by Nielsen Norman Group usability heuristics — reimagined for agentic systems, AI agents, and multi-agent environments.
Visibility of Agent Intent & Action
Users shouldn't guess why the agent is doing something. Use patterns like:
- Intent playback
- Plan preview
- "Here's what I'm about to do…"
Progressive Transparency (Not Full Transparency)
- Early: full reasoning, sources, confidence
- Later: concise or silent execution
This is core to AX: Transparency → Summary → Confident → Silent
User Control Through Steering, Not Micromanagement
Replace “Click here → do this → confirm that” with:
- Approve / modify / redirect
Key shift: UX control → AUX steering
Trust is Dynamic, Not Static
- New user: cautious, gated
- Experienced user: autonomous, fast
Modeled via trust timeline and progressive stages.
Clear Boundaries of Autonomy
At any moment, users should know what is:
- Autonomous — agent acts freely
- Confirmed — agent asks first
- Blocked — agent cannot act
Graceful Handling of Uncertainty & Failure
Instead of hallucinating:
- Ask
- Escalate
- Provide partial output
Includes confidence thresholds, escalation paths, and failure design as first-class.
Appropriate Agent Assertiveness
Assertiveness levels:
- Compliant → Advisory → Assertive → Protective
“Pushback is professional.”
Context Efficiency & Awareness
Avoid full scans and context overloading. Prefer:
- Routing
- Indexing
- Summaries
Core AUX problem: context = cost + performance constraint.
Multi-Actor & Multi-Agent Clarity
Human vs agent vs agent. In multi-agent systems:
- Who delegated?
- Who owns the outcome?
Consistency of Behavior, Not Interface
Unlike classic UX, the UI can adapt (generative UI). But these must remain stable:
- Decision logic
- Trust boundaries
- Tone
Apply these heuristics
to your product
We help teams design agentic experiences that people actually trust. Let's talk about your next build.
Start a Project →