Your design system is now AI infrastructure

Design systems were sold as a way to keep designers and engineers consistent. That framing is outdated. In 2026, the primary consumer of your design system might not be a person at all. It's the codegen tool, MCP server, or agent drafting components from Figma files while your engineers sleep.

Tokens become prompts

LLMs generate reliable output when input is structured and unambiguous. A token named `color-text-muted` maps cleanly to intent; `grey3-final-v2` does not. Every ambiguous variable name, inconsistent spacing scale, and undocumented component variant is a place where automated pipelines will guess, and sometimes guess wrong at scale.

What changes in practice

Semantic naming stops being a nicety and becomes a correctness requirement. Component APIs need to be regular enough that an LLM can infer usage from context. Documentation needs to be written for two audiences: humans skimming, and agents parsing. None of this makes design systems worse for people. The discipline that helps machines usually helps new hires too.

The audit question

If you're investing in AI-assisted development, audit your design system's machine-readability before blaming the models. We've seen 'AI produces bad code' turn out to be 'the design system encodes six different ways to build a card.' Fix the source of truth and the pipeline improves overnight.