Product DesignAI

Design systems are becoming agent systems

Jul 2026 · 7 min · Draft

A design system used to speak to people. Its next reader is a machine — and that quietly changes what a design system is.

What is a design system, reduced to its essence? A set of promises. This button will always look like this. This spacing always means that. Break the promise, break the product.

Promises

Tokens, components, and guidelines are all the same thing wearing different clothes: promises about future decisions, made once so they don’t have to be re-made a thousand times. For twenty years those promises were kept by humans, through discipline — reviews, audits, and the slow socialization of “how we do things here.” Discipline is expensive, and it decays. Every mature system drifts.

Readers

Now the system has a new audience. Agents are building interfaces — scaffolding pages, composing components, writing the CSS your juniors used to write. And documentation written for human judgment fails machine readers in a specific way: ambiguity. A senior designer routes around a vague guideline without noticing. An agent turns the same vagueness into a defect, at scale, in seconds. Writing for machines forces the honesty our documentation always claimed to have.

Fig. 01 — Two readers, one system: what the human infers, the machine must be told.

Executable

Here is the shift in one sentence: documentation becomes executable. A rule that can be checked is a test. A token that can be served is an API. A guideline that can be enforced at generation time never needed to be a PDF. The design system stops being a website you visit and becomes infrastructure you run — shipped in code, versioned like code, consumed at build time by whoever, or whatever, is doing the building.

A rule that can be checked is a test. A token that can be served is an API.

Memory

Systems drift because memory lives in people, and people leave. The why behind a token — why 22 pixels, why this gray and not that one — evaporates, and with it the ability to evolve the system without breaking it. Agent systems need that memory explicitly: decisions as first-class objects, stored with their reasons, served to every agent at the moment of work. This is what I built Engrym around. The design system becomes the agents’ memory — and, for the first time, the organization’s.

Fig. 02 — Decision memory: the why, kept, versioned, and served.

Convergence

Follow both threads far enough and they meet. The design system and the agent orchestra converge into a single artifact: a machine-readable definition of what good means here. Teams that have one will ship coherent products at a speed that looks unfair. Teams that don’t will ship faster too — incoherently. The gap between design and AI isn’t bridged by tools. It’s bridged by systems both can read.