A bounded pilot with an AI agent and your design system

The ranking reads a design system from its makers’ pages. A pilot meets the working parts. Does the server install in your setup? Do the docs answer the questions your screens raise? Does the guidance hold when the prompt gets awkward? Small faults only show up in use. GitHub Primer’s MCP server, for one, had 2 tools that returned the home page instead of the right component for names of more than one word, until a fix in October 2026. A pilot finds that kind of fault on 3 screens, before the whole product depends on the server.

Write these down, and do not change them once the pilot begins:

• Scope. 2 or 3 named screens your team needs anyway, one of them awkward.
• Setup. The agent, the model, the server and the rules files, frozen for the length of the pilot.
• End date. A day on the calendar.
• Acceptance. The checks each screen must pass, and the person who signs it off. Checking an agent’s output sets them out.
• Baseline. How long the same kind of screen takes your team today, measured from recent work rather than guessed.
• Decision rule. What result means adopt, what means fix first, and what means stop.

Start by putting your test set of prompts for AI agents through the agent, so you have a first score. Then build the screens. Every time a person corrects the agent’s work, log it in one line: what was wrong, and which part of the system should have prevented it. Fix rules and design tokens between sessions, never in the middle of a screen, and rerun the test set after each fix so you can see whether it helped.

Sort the correction log by cause. Each cause points to its own fix:

• A value the agent invented. A design token is missing, or not described well enough to find. See design tokens as data.
• A component it rebuilt. The catalog was hard to search, or not in reach. See agent access.
• A rule it broke. The rule was not written down, or not written for agents. See written guidance.
• A defect in the system itself. Report it to its makers, or fix it in your own.

A short log that keeps shrinking means adopt. A long one where every line has a fix means fix first and run again. And when the system works but the output still does not look like your product, stop, because that is as far as the public route goes.

A public system can pass every check and still wear its maker’s look. When a pilot ends there, the next step is a custom AI-ready design system built around your brand, with the same kind of checks. That is the half of the category the best AI-ready design systems, ranked cannot cover, and it is what Work with Simbolik (opens in a new tab) offers; bring the correction log to the call.