Skip to content
Article Substack Mar 2026

Vallaure: five changes design systems need before agents can use them

What the article is about

Christine Vallaure’s article on Substack addresses a concrete problem: most design systems were built to be used by human designers, which means they carry assumptions that AI agents cannot easily interpret. The article proposes five specific structural changes that make a design system more legible to AI agents working inside Figma or similar tools.

What it covers

The first change concerns tokens. Vallaure argues that semantic tokens — tokens named by their intent (“primary-action”) rather than their value (“blue-500”) — give agents enough context to apply the right styles without guessing. A token named by its role can be matched against a component’s function; a token named by its hex code cannot.

The second change is about component architecture. Props in Figma components should match the props that corresponding code components expose. When the two sets diverge, agents face a mapping problem: a button variant in Figma and the same button variant in the codebase describe different things, and agents operating across both layers will produce inconsistent output.

Third, Vallaure calls for documenting all component states, including disabled, error, loading, and edge-case states. Agents filling in a form template, for example, need to know what the empty and filled states look like to apply them correctly.

The fourth change is structural: use auto layout throughout components, with spacing values drawn from tokens rather than fixed numbers. This allows an agent to resize and rearrange elements while staying within the spacing system.

The fifth change is straightforward: add plain-language descriptions to components. Agents can read these descriptions to understand what a component is for and when to use it, rather than inferring purpose from visual appearance alone.

What it does not cover

Vallaure acknowledges three areas where current design systems remain poorly equipped. There is no established workflow for an agent to propose a design change and have a human designer review it before it is applied. Visual brand attributes — tone, personality, aesthetic choices that are not reducible to tokens — are not capturable in current token structures. And governance at scale, the question of how to manage agent-generated design contributions across a large organization with many contributors, has no proposed solution yet.

Who it is useful for

Design system leads and senior designers working with teams that are experimenting with AI agents in Figma. The article is practical rather than speculative: it names changes that can be made now, not changes that require AI tools that do not yet exist.