Designing for Agentic AI: UX Patterns for Control, Consent, and Accountability#
UX design patterns guide (IDE, February 2026) for building trustworthy agentic systems. Provides the human-agent interaction framework missing from the wiki’s architectural analysis. Only 6% of companies fully trust AI agents for core processes.
Core Thesis#
Designing for agentic AI is designing for a relationship. Autonomy is an output of a technical system; trustworthiness is an output of a design process. The shift from UX to AX (Agent Experience).
Six UX Patterns#
| Pattern | Phase | Purpose | Key Metric |
|---|---|---|---|
| Intent Preview | Pre-action | “Here’s what I’ll do. OK?” | >85% acceptance |
| Autonomy Dial | Pre-action | User sets independence level per task | Setting churn |
| Explainable Rationale | In-action | Proactive “why” grounded in user prefs | “Why?” ticket volume |
| Confidence Signal | In-action | Agent communicates certainty level | Calibration score >0.8 |
| Action Audit & Undo | Post-action | Timeline + reversal for every action | <5% reversion rate |
| Escalation Pathway | Post-action | Agent asks for help vs guessing | >90% recovery success |
Autonomy Dial Levels#
- Observe & Suggest — notify only
- Plan & Propose — user reviews every plan
- Act with Confirmation — final go/no-go
- Act Autonomously — pre-approved tasks, notify after
Maps directly to claude-code’s 6 permission modes and the wiki’s autonomy spectrum (Theme 3).
Phased Adoption#
Phase 1: Intent Preview + Audit/Undo (foundational safety) Phase 2: Autonomy Dial + Explainable Rationale (calibrated autonomy) Phase 3: Act Autonomously for pre-approved tasks (proactive delegation)
Service Recovery Paradox#
A well-handled mistake can build more trust than flawless execution. Acknowledge error → state correction → provide human support path.