AI Native Design System — case study cover

Élan Design System is a living example of human-AI co-evolution and harnessing. Unlike others, Élan's token architecture and documentation are designed convenient for an AI agent to read, generate, and self-correct with minimum human intervention.

Agent Harness Architecture

The harness that constrains and co-evolves with the AI agent — dual evaluation, 3-domain feedback routing, and guardrails that tighten with every repeated failure.

harness constrainsknowledge informs
Drag to pan · Scroll to navigate · Click a node for details
View the full design system →

Dual evaluation (AI pre-check + human review) · 3-domain routing · Self-tightening guardrails


Agent-Native Semantic Tokens

A token naming convention designed for machine readability — every name fully describes its intent, so agents can generate and validate tokens without a lookup table.

Naming Formula

Pick a color swatch to begin — the role auto-fills and available dimensions filter to valid combinations.

.property.role.emphasis
Property

What element type the color applies to

Emphasis

Prominence level within its group

Why Semantic Naming
IBM CarbonFlat referential: $carbon--blue-60Encodes hue + grade but not purpose — an agent must map blue-60 → brand vs. informational from external context
Material DesignRole-based: --md-sys-color-primarySemantic but shallow — 'primary' doesn't distinguish surface vs. text vs. icon usage
Goldman Sachs (One GS)Property-first: property · role · emphasisDeep semantic structure — each dimension independently queryable

Adopted the Goldman Sachs property · role · emphasis hierarchy because each token name fully describes its intent. An AI agent can parse color.surface.brand.bold without a lookup table — the name IS the documentation.

Explore the full color system in the playground →

Goldman Sachs–inspired property · role · emphasis naming · Machine-readable by default


Systemic Pattern Map

Patterns invisible to individual incidents become obvious in aggregate. Four categories hit the 3-strike threshold and promoted to hard guardrails.

guardrailed audited documented
Explore the full design system →

47 incidents · 6 failure categories · Documented → Audited → Guardrailed


ScrollSpy — A Micro-Interaction Deep Dive

One component, 7 anti-patterns, 4 iterations. Click and drag share the same pointer — a 3px dead zone discriminates intent.

Before
Intro
Tokens
Components
Process
Result
Hover over the rail
Every tap triggers a drag — no click detection
After
Intro
Tokens
Components
Process
Result
Tap or drag the rail
< 3px = click, > 3px = drag — intent detected
Try the ScrollSpy in the playground →

7 anti-patterns · AP-011 pointer mapping · AP-012 dead zone · AP-031 Framer Motion transform conflict