What does a checkable design system line look like?

Reuse the color tokens becomes: no hard-coded color value in a diff. Broken when a hex code appears outside the stylesheet. Checked by a lint on every pull request. The rule, the moment it broke, and what catches it.

What about rules a machine cannot check, like tone?

Name the reviewer and the moment in the process when they look. A sample is fine. A line that needs judgment and has no named person is a principle, and a principle has no failure condition.

Where does the Figma agent fit?

On the check side. Per Wayne Lin's October 6, 2026 post it can flag a custom button that duplicates the system's Button/Secondary component or uses a hard-coded color. It runs when a designer asks, and a team can package the check as a skill to run it regularly.

The design systems team. In Design's Kill List I replace the component library as monument with a constraint layer and a much smaller set of primitives, with the systems team owning the constraints and the checks.

How to Write a Design System for AI Agents

How do you write a design system for AI agents?

THE SHORT ANSWER

Write each line as a rule with a failure condition and a check, and sort the lines into three piles. Luke Wroblewski's October 1, 2026 post describes the reading side: an AGENTS.md file that tells agents to reuse the site's tokens and components, and a design system page built from the same code. Reading is not following. Vercel's inbound sales agent read a 1,000-line prompt on every run and stopped following some of its rules, per Jeanne DeWitt Grosser in Tomasz Tunguz's write-up, so 14 rules moved into code. In my Design Operating Model a constraint is a rule with a failure condition attached. Lines a machine can catch go in the pipeline, lines that need judgment get a named reviewer, and lines with neither get cut.

Give every line a failure condition and something that catches it.

Luke Wroblewski's post on design systems for AI agents describes the reading half well. On his team's Intent site an AGENTS.md file tells agents to reuse the theme classes and CSS variables in the global stylesheet, and the design system page is built from the same code as the site, so it stays current. His closing line is that agents read the documentation every single time.

They do. Vercel's inbound sales agent read its prompt every time too. Jeanne DeWitt Grosser told Tomasz Tunguz the prompt grew to 1,000 lines and the model would not always follow some of the rules in it. Fourteen of them moved into code.

What I see in teams copying the idea is an agent-facing design system written in prose: use the brand voice, prefer existing components, keep spacing consistent. Rule 1 of my Design Operating Model, in Design Just Got Promoted, is that a principle has no failure condition and a constraint is a rule with one attached. Its worksheet pairs every rule with a failure condition and a check.

So sort the file into three piles. Lines a machine can catch get rewritten with the failure condition and go in the pipeline, the way "reuse the color tokens" becomes a lint for hex codes outside the stylesheet. Lines a person has to judge get a named reviewer and a moment. Lines with neither get cut, because every unchecked line dilutes the checked ones.

Then count how many lines have a checker. That number is the state of your design system for agents.

The full piece is Agents Read the Design System. Read Is Not Checked.

Last reviewed 2026-10-07 · 2 min read

How do you write a design system for AI agents?

SOURCES

THE LONG VERSION

RELATED ANSWERS