Machine-readable logic

Your design system has a new user, and it cannot see. An AI agent now builds interfaces from your system, but it cannot look at your Figma file. It can only read. So the job flips: from documenting for people to encoding for machines. If the system is not written for a machine, the agent guesses, and guesses drift.

Tokens told the agent what. This layer tells it when, why, and when not. The work calls it an agentic design system: infrastructure an agent can read, reason over, and build with, through encoded intent and constraints.

Here is the failure everyone hits, the Crimson Trap. An agent needs a color for a Delete button. It finds brand-crimson-500 and system-red-600. Both red. It picks the brand one. Valid, and wrong, because only the second means danger. The agent did not slip. You never told it which red means danger. Skip the meaning layer and the model falls back on the average of the internet's UI, not yours.

The proof is measured, not argued. Diana Wolosin, then at Indeed, with Tony Rucker building the MCP infrastructure, ran a 1,056-prompt benchmark across eight ways of feeding a system to an agent. Their human-written docs were verbose, around 30,000 tokens a query, and structured JSON beat them on both cost and accuracy, roughly five times cheaper. Read the lesson correctly, though: it is structure, not size. Diana's actual rule is JSON for MCP, Markdown for LLM. Structured data like props, sizes, and variants belongs in JSON because it is a contract; natural-language rules still belong in Markdown, with front matter instead of bloat. And it is one team's numbers on one system, so treat it as direction, not law.

Structure is necessary, not sufficient, and this is the honest part. When Indeed put the winning format into production, the team generated 4,300 prototypes in four months, and an audit of a sample still found typography violations, broken spacing, and an invented color palette nobody approved. Structured context raises the floor. It does not remove the need for governance and human review on top. Governed well, the upside is real: Jesse Gardner at New York State turned a five-page state PDF into a working, accessible form in about thirteen minutes, with the system feeding the agent real component code instead of guesses.

You do not rebuild everything to get there. You annotate what you already have: label your top tokens with intent, declare a naming grammar, say in Figma when not to use a component, and put always-on rules in one AGENTS.md.

The teams getting real leverage from AI are not the ones with the smartest agents. They are the ones whose design system the agent can actually read. And once the agent can read the meaning, the next question is what it builds against. Not ad hoc prompts, but a written, versioned spec. That is the intent layer.