Medium: Building agentic design systems

Luis Ouriach, a designer who works on design systems at scale, wrote this piece in December 2025 as a framework for how design teams should structure their relationship with AI tooling. The article is not a tutorial. It addresses a more fundamental question: why isolated AI experiments fail inside design organizations, and what a coherent approach looks like.

The central argument is that a design system must expand beyond its traditional scope. A library of reusable UI components is one node inside a larger, interconnected brand operation that now includes email templates, marketing websites, and brand voice. Ouriach calls the expanded form an “agentic design system” — one enhanced with AI capabilities to automate repeatable tasks while preserving human judgment over strategic direction.

Ouriach identifies fragmentation as the main obstacle most teams face. Designers now routinely use Cursor, Replit, or Claude in personal workflows, but the outputs stay local. When each person builds AI habits in isolation, the organization accumulates scattered experiments rather than shared knowledge. His recommendation is direct: before anyone opens an AI tool on a task, the team writes a brief together. The prompt comes after the conversation, not instead of it.

The article also addresses a specific technical question about framework choice. Because most AI code-generation models were trained on large amounts of open-source code, frameworks with broader representation in that training data produce more reliable outputs. Tailwind CSS has become the preferred choice for design-to-code work precisely because tools trained on Tailwind-heavy codebases generate cleaner, more consistent results than tools working with bespoke CSS architectures.

On the question of what AI can and cannot do, Ouriach is clear. AI handles speed and repetition. It cannot replace a well-formed brief or an exploratory design session that stretches brand conventions in new directions. Teams that deploy AI before establishing clear design principles end up producing consistent output in the wrong direction.

Two practical recommendations run through the piece. First, maintain critique cultures even when AI accelerates early production — the ability to evaluate and reject output matters more, not less, when output volume increases. Second, invest in documentation of custom frameworks and codebases, because AI tools operate on the context they receive, and teams that describe their systems thoroughly get better results than teams that expect the tool to infer their conventions.

This article is most useful for design leads and design system owners at organizations with more than one designer, where the question is not whether to adopt AI, but how to integrate it without losing collective quality control.