We handed an AI agent a tokenized style guide and asked it to build production UI. It got the design 80 to 90 percent right. This essay is about the missing piece, because the missing piece is the whole argument for machine-readable design systems.
Here's the setup. I run an enterprise design system at a Fortune 250 utility, and we ship an agentic platform to about 8,000 internal users. We built one portion of it on a tokenized style guide: real tokens, decent documentation, no enforced JSON contract behind any of it. The style guide was written for humans who could fill gaps with judgment. The agent had no judgment to fill gaps with. It had the guide, and it had the average of the internet.
What held up surprised me. The page level came through clean: overall layout, structure, navigation, the relationships between screens. If you squinted, it looked shipped.
Then you stopped squinting. The slippage was all at the element and component level. A prop that didn't exist. A variant that almost matched. A spacing value that was close, which in a design system is another word for wrong. The granular details inside individual components, the exact places where "roughly right" doesn't count.
That distribution of errors is not a coincidence, and it's the most useful thing the experiment taught us. Page-level structure is forgiving. There are many acceptable ways to arrange a screen, so a probabilistic system lands somewhere acceptable. Component-level decisions are not forgiving. There is one correct token for that button label. The props are the props. A component contract is binary: you match it or you don't. Asking a probabilistic system to hit a binary target without a contract is asking it to guess, politely.
Which means the 80 percent wasn't the agent being impressive. It was the agent consuming the part of our system that was already structured, then improvising the part that wasn't. The style guide's tokens gave it real values to work with. Nothing gave it the rules for applying them. Values without contracts get you exactly what we got: the shape of your product with someone else's details.
The industry data says our experience is the norm, not the exception. Indeed's benchmark work (Diana Wolosin, with Tony Rucker building the MCP infrastructure) found that structure, not volume, is what moves agent accuracy: their structured component data beat verbose documentation on both cost and correctness. And when a winning format did reach production scale, a team generated 4,300 prototypes in four months and an audit still found typography violations and an invented color palette. Structure raises the floor. Governance holds the ceiling. Nobody skips either.
So the honest reading of "the agent got it 80 percent right" is: the agent revealed exactly where our system stops being infrastructure and starts being vibes. That's a gift, if you treat it as one. The gap map is the roadmap. Ours said: the component contract layer is next, and it has to be enforced, not suggested. A schema the agent cannot misread. Props, variants, and the specific token applied, expressed as data instead of prose.
Two things I'd tell any team about to run the same experiment.
First, run it before you feel ready. The 80 percent build took us days, not months, and it produced the most precise backlog we've ever had. You are not risking production; you're prototyping your system's legibility.
Second, resist both wrong conclusions. "The agent failed" is wrong: it performed exactly as well as the structure we fed it allowed. "The agent basically works, ship it" is wrong too: the last 20 percent is where users live. The right conclusion is quieter. Eighty percent from a loose style guide is the floor. The whole point of the tokens-to-agents stack (tokens, machine-readable meaning, specs, runtime context) is to raise that floor toward something you'd put your name on.
The agent will tell you the truth about your design system faster than any audit you'd commission. Ask it.
Want the checklist version? The free agent-ready starter kit at tokenstoagents.ai/kit includes the readiness checklist we now use before any agent touches UI.