Does a design system change what a coding agent writes?

Mateusz, an engineering manager, had an AI agent build one app 30 times and found that an installed design system stopped it rewriting components, while written guidelines did not.

Mateusz, an engineering manager in Poland, built a design system for Angular, a web framework, with 14 components published as installable packages and guides written for AI agents. He wanted evidence that it changes what an AI coding agent writes. He had an agent build the same small dashboard 30 times, changing only how much of the design system it could use.

• Without the installed package, the agent wrote its own code in 12 or 13 of the 13 places where a library component fitted, and added about 90 custom style values. With the package, it did neither.
• In 30 runs the agent never invented a component that did not exist; it wrote its own instead, which is harder to notice in review.
• The agent copied the colours correctly from the Figma design file but repeatedly produced text that failed contrast checks, on both AI models tested.
• Using the library cost more per run, because everything the agent read stayed in its working context and was paid for again on every turn.

Mateusz concludes that installing a design system decides whether the application or the system ends up owning the code an AI agent writes. The study covers one application, one framework and one AI model vendor, and he wrote both the design system and the evaluation.

Derived by Working Surface from the article; more in the Playbook. Source line: The most useful thing the system does turned out to be accessibility, not consistency.

14 October 2026: as AI agents raise the volume of code changes, teams are moving routine checks into automated tests and installable rules, and keeping people for intent and contested decisions.