The Problem

As AI agents move from chatbots to autonomous decision-makers, traditional UX patterns break down. Designers ship agents without structured frameworks, leading to opaque black boxes, permission fatigue, silent failures, and runaway autonomy.

Existing UX heuristics (Nielsen, et al.) were built for deterministic interfaces — a click does the same thing every time. They don’t translate cleanly to probabilistic systems that interpret intent, take consequential actions, and fail in unfamiliar ways.

The framework is my attempt to fill that gap with something practitioners can actually apply.

The Framework — Five Layers, Twenty-Five Principles

Layer 1 — Intent. Users say what they want, not how to do it. How agents interpret goals, confirm understanding, and remember preferences. Intent over instruction · Confirm before commit · Graceful ambiguity · Progressive disclosure · Intent memory.

Layer 2 — Trust. Trust is earned, not assumed. How agents earn confidence through transparency, uncertainty honesty, and reversibility. Show your work · Communicate uncertainty · Make rollback effortless · Earn autonomy incrementally · Fail transparently.

Layer 3 — Guardrail. Agents need limits. Spending caps, scope boundaries, escalation paths, audit trails. Explicit boundaries · Cost awareness · Blast radius containment · Human escalation · Audit trail.

Layer 4 — Orchestration. Agents rarely act alone. Coordination patterns for multi-agent systems — handoffs, conflict resolution, shared memory, role clarity, failover.

Layer 5 — Evolution. Agents improve over time. How systems learn responsibly: feedback loops, behaviour versioning, deprecation paths, user-controlled adaptation.

The Agent UX Framework — User Journey: 'Book me a flight to Berlin' flowing through Intent, Trust, Guardrail, Orchestration, and Evolution layers into a trusted action
The user journey through the five layers — intent becomes a trusted action only after passing every checkpoint. Each layer is a checkpoint, not a checkbox.

Why It Matters

Each principle maps to a concrete design decision: a UI pattern, a confirmation flow, a permission model, a failure mode. The framework bridges the gap between high-level “AI ethics” abstraction and the actual screens, flows, and system prompts that ship to users.

Abstraction without specificity is decoration. The framework’s whole point is that every layer ships with canvases designers can fill in and patterns engineers can implement.

Status

v1.0 published March 2026. Currently iterating in public — new patterns added as agent products land in the wild and either reinforce or break a layer’s heuristic. Companion canvases (one per layer) in early build; release path TBD.

What’s Coming