Three design systems that have opened up to agents

“AI-ready” now appears on a lot of slides. A more reliable approach is to look at teams that have publicly described what they built and what they measured. The following three examples show what they prove and what they leave open.

Indeed: documentation as data

Diana Wolosin of the design system team at Indeed converted the documentation of 77 components from MDX files into structured metadata and tested eight MCP configurations against each other with 1,056 prompts. According to the summary at Into Design Systems, the Markdown version needed around 30,000 tokens per request, covered 82 percent, and hallucinated along the way. The structured JSON version used about 80 percent fewer tokens and was more accurate.

Two caveats apply. I'm relying on the summary of a talk here rather than on the raw data. In addition, cost does not work as an argument, since the report puts it at 300 versus 1,500 dollars a year. The gain is in accuracy.

Atlassian: a dedicated MCP server

The Atlassian Design System team built its own MCP server, plus structured content files that feed agents, prototyping tools, and also the atlassian.design website. By its own account, for tasks in the design system, accuracy rose by 52 percent and speed by 34 percent on average, while tool calls and token usage fell by 26 and 16 percent. The post doesn't say what baseline this was measured against.

These are the vendor's figures about its own system. They indicate a direction but are unsuitable as a benchmark for other systems.

IBM: querying Carbon via MCP

IBM offers an MCP server for Carbon that agents can use to query components, tokens, icons, guidelines, and code examples. Anyone who doesn't work at IBM currently has to request access. I didn't find any measurements for it.

What the three have in common

All three make their design system knowledge queryable for agents instead of relying on prose. None of the examples proves that this has the same effect in every system. What is proven is that the approach was implemented and, where it was measured, paid off.

For your own system, only a measurement of your own provides clarity. This means running the same tasks with and without a structured description and counting how often the agent makes something up.

From the series “Design Systems and AI”. Previous: Why coding agents fail at the design system. Next: Contracts instead of essays: describing components in a machine-readable way.