ai-native

AI-native design specifies interfaces whose screen, affordances, and content are produced at runtime by a model rather than authored ahead of time — agent UX, conversational and agentic UI, and generative non-deterministic UI. There is no settled canon for this pillar. Unlike the other design pillars (each anchored in established books), this one is principle-derived: drawn from practitioner essays and emerging sources, marked as such on every claim. Cite the principle and the canon gap.

When this applies

• Designing an agent UX / AX surface — a system that takes goals and acts, rather than rendering controls the user operates step by step.
• Designing conversational or agentic UI — chat, tool-calling assistants, multi-step autonomous flows.
• Designing generative / non-deterministic UI — where the model assembles the layout, copy, or component set at request time, so no two sessions are identical.
• Building a no-fixed-screen interface — no stable set of pages to map, no fixed affordances to learn.
• Calibrating trust and control for AI output — confidence display, undo, correction, escalation to a human or a deterministic path.
• Diagnosing which classical pillar strains when AI removes its fixed substrate, and what to substitute.

Not for: whether a conventional screen is operable — heuristic evaluation, UX laws, affordances on a fixed UI (use usability); the visual DNA, fonts, or color (use core design); a voice or conversational surface treated as a device class with its own layout constraints (use core surface).

Rules

• No settled canon — say so. This pillar has no book-anchored canon (research: design-pillar-landscape, "🔵 emerging, no canon"). Every recommendation names its principle-level source (Dibia agent-UX principles; Wilson Age of Invisible Machines 2022; Smashing "Designing for AI Beyond Conversational" 2024) and flags that it is principle-derived, not established practice. Do not present it as settled. Load references/ai-native-caveats.md before making strong claims.
• Don't over-claim. When the right answer is "this is unsettled — here are the trade-offs and the current thinking," say that. Confident prescription on an unsettled question is the failure mode this pillar exists to avoid.
• AI breaks the frameworks — name which one, then substitute. The classical canon assumes human-authored, screen-delivered, click-based UIs (research COUNTER #2). When AI removes a fixed substrate, a specific pillar strains. Diagnose which, then bring the substitute (the Procedure). Don't apply a strained framework as if it still held.
• Cite usability down where it still applies. Inside an agent UI, conventional operability still governs the parts that are fixed — the input box, the controls, the confirmation step. Cite usability laws (Fitts, Hick, Nielsen) by name for those; never re-derive them here. Citation points down to usability, never the reverse.
• Stay generative. Produce the design recommendation or the diagnostic — don't degrade into a taxonomy of what's uncertain. The framework-strain model is the organizing spine, not a box-ticking sequence.

Procedure

A. Orienting question: is the substrate fixed or generated?

Load references/ai-native-principles.md. Establish what the AI removes:

The interface has…	Then…	Reach for
A fixed set of screens/pages	Journey mapping still works	core/journey — not this pillar
Generated layout/content per request	No stable pages to map	the generative-UI section below
Goal-taking agent behavior	No fixed click-path	the agent-UX section below
A conversational front door	Turn-taking, repair, grounding	the conversational section below

B. The framework-strain diagnostic

When AI removes a fixed substrate, name which classical pillar strains and substitute (research COUNTER #2; depth in ai-native-principles.md):

AI removes	Classical pillar that strains	Substitute the pillar with
Fixed screens	Journey mapping (journey) assumes a stable touchpoint sequence	Intent-and-outcome framing: design the goal contract and the recovery paths, not the page sequence
Stable affordances	Interaction design (signifiers/affordances, Norman) assumes the user learns a fixed control set	Progressive disclosure of capability; make the agent's actions inspectable and reversible
Stable content	Content design (content-design) assumes authored copy	Voice/tone as a generation constraint the model obeys, plus guardrails on what it may say
Autonomous action	Ethics (deceptive-patterns) assumes a human pulls each lever	Consent, transparency, and a human-in-the-loop checkpoint before consequential action

C. Agent UX / AX

When the system takes goals and acts:

1. Design the goal contract — what the user asks for, what the agent will and won't do, where it stops to confirm.
2. Make the agent inspectable — surface intermediate steps, tool calls, and sources so the user can audit, not just trust.
3. Make consequential actions reversible or confirmed — undo, or a checkpoint before anything destructive (cite usability: error prevention, Nielsen #5).
4. Design for failure and handoff — what happens when the agent is wrong, low-confidence, or out of scope; the escalation to a human or a deterministic path.

D. Generative / non-deterministic UI

When the model assembles the interface at runtime:

1. Constrain the generation space — a component vocabulary and layout grammar the model composes within, so output is novel but not chaotic.
2. Keep fixed anchors — the parts users must relearn nothing about (primary input, navigation home, the stop control) stay stable across sessions; cite usability for those.
3. Plan for non-determinism — the same request may render differently; design so that variance never breaks trust or task completion.

E. Conversational / agentic UI

When a conversation is the front door:

1. Design turn-taking and repair — how the user corrects a misunderstanding, interrupts, or restarts.
2. Provide grounding — show what the system understood before it acts on it.
3. Don't make conversation the only path when a fixed control is faster (cite usability: a button beats a sentence for a known action).

F. Trust and control

Across all of the above:

1. Calibrate confidence — show uncertainty honestly; never present a guess as a fact (mirrors the core cite-the-principle / no-unsourced-opinion posture).
2. Always leave an exit — a way back to a deterministic, human-operable path.
3. Mark the canon gap in the deliverable — note where the recommendation is principle-derived and unsettled, so the reader weights it correctly.

References

Reference file	Load when
references/ai-native-principles.md	The constructive layer: the framework-strain model in depth, agent-UX principles (Dibia/Wilson/Smashing), generative-UI patterns, conversational/agentic UX, trust-calibration techniques.
references/ai-native-caveats.md	The honest-citation layer: the no-settled-canon warning in depth, the over-claim guard, where conventional operability (usability) still governs, and how to weight principle-derived guidance. Load before any strong or prescriptive claim.