Your Design System Is Now an API for Coding Agents

As AI increases the amount of software that teams can produce, designers and developers will not be able to participate in every implementation decision. Their highest-leverage work increasingly becomes defining the systems, constraints and evaluation loops that allow others, including agents, to build reliably.

The design system as an agent interface

Meta's Astryx is perhaps the clearest current example of this direction. Astryx is an open-source React design system derived from an internal system used across more than 13,000 Meta products. Its components, documentation, theming system and CLI were explicitly designed so humans and AI assistants work from the same reference.

A CLI lets agents search the system, inspect components and retrieve examples rather than loading the entire documentation into the context window. The same knowledge can be consumed by developers, tools and agents.

Astryx goes a step further and runs evals to test whether the documentation and tools actually make the system easier for agents to use. That points toward another important change.

From design review to design evals

When a designer can review every important screen, quality can live partly in that person's judgement. But with agents producing hundreds of changes across a product, manual review does not scale at the same rate as generation. Some of that judgement has to move into the system.

The designer's role moves toward the system

This does not mean designers stop designing screens. It means designing each screen manually becomes a smaller share of where leverage can come from. When interface generation becomes cheap, one of the most valuable things a designer can do is improve the environment in which interfaces are generated. In that sense, designers need to compile their judgement into the product-development system.

The work is not one output, but hundreds of future outputs.

Design systems in the AI age

The first generation of design systems helped designers and developers build consistent products faster. The next generation will have another responsibility. They will need to make an organization's design knowledge understandable to machines. That requires more than a component library. It requires machine-readable knowledge, canonical examples, constraints and constant evaluation.

As AI makes producing another interface increasingly cheap, the scarce work moves elsewhere. It moves into deciding what should exist. Defining what good looks like. Encoding that judgement into systems. And building feedback loops that allow increasingly autonomous agents to operate within predefined boundaries.

The design system stops being only a tool for implementing the product. It becomes part of the infrastructure for producing the product.