How to choose a design system for AI agents

A design system is chosen for a product, and an agent only inherits that choice. So before you look at any system, write down which agents your team runs, what they make (admin screens, marketing pages, a mobile app) and in which stack. Most public systems are tied to one. Ant Design is React for data-heavy tools. Shopify Polaris is for apps inside Shopify. Salesforce Lightning is for Lightning Web Components. The GOV.UK Design System is for UK government services. A system outside your stack is not a candidate, whatever its score.

With the shortlist made, read each system’s 4 ratings on the best AI-ready design systems, ranked and ask what each gap would cost your team.

• No design tokens as data. The agent copies values from CSS or guesses them, and you export the design tokens yourself.
• No official server. The agent works from files alone. That can be enough when the docs are good: Microsoft Fluent 2 has no public server and still scores 5.
• No docs as text. The agent reads web pages built for people, or falls back on what it remembers from training.
• No guidance for agents. You write the rules yourself, before the first prompt.

A gap you are ready to fill is cheaper than a system that does not fit.

A Yes says the part exists on the makers’ own pages. It does not say it comes free of conditions, and some of those decide a choice:

• Carbon’s hosted MCP server asks for an IBMid sign-in.
• Atlassian’s skills for agents are marked private to Atlassian; its CLI, server and full-text docs are public.
• Salesforce Lightning’s agent tools sat inside Salesforce’s general MCP server until Salesforce retired it on November 2, 2026, so an agent now reaches SLDS through skills alone.
• Chakra UI’s server has tools for its paid templates that need a key.
• Shopify Polaris’s React package is deprecated in favor of its web components, and its server covers all of Shopify, not Polaris alone.
• Material 3 for the web is in maintenance mode, according to its own repository.

Open the pages behind your shortlist’s ratings and look for the same kind of line.

Three routes come out of the shortlist. Adopt a public system that scores 7 or 8 and fits your stack, and accept its look; your agents get the most help for the least work. Adapt one by theming it; the agent keeps the maker’s docs and server, but they describe the maker’s defaults, so every change you make needs guidance of its own. Build your own when the output must look like your company and no one else; it can be AI-ready from the first file. The cost and scope of making a design system AI-ready sizes that route, and Work with Simbolik (opens in a new tab) is one way to have it built.

A score is read from documentation. Your agents will meet the real thing. Put the front runner through a bounded pilot with an AI agent on 2 or 3 real screens before anyone migrates a page.