What Astryx's AI-agent manifest gets right that most design systems don't
Meta's newly open-sourced design system ships a JSON contract built for coding agents, not just humans. It is a fix for a bug we have hit ourselves.
AI agents hallucinate props from prose docs. Astryx fixes that, and it is a pattern worth stealing.
We have watched an AI agent hallucinate a component prop that does not exist, three times this month, across three different UI libraries. Meta's newly open-sourced Astryx design system ships a fix for exactly that problem, and it is a pattern more dev tools should copy.
The actual bug
Point an agent like Claude Code at a UI library to scaffold a screen, and it reads whatever documentation exists, a README, a Storybook page, a markdown props table, and infers the API from prose. Inference from prose is lossy. The agent invents a prop name that is close but wrong, assumes a component takes children when it actually wants a named slot prop, or reaches for a deprecated variant because an old blog post happened to rank higher in its context. We fix this by hand, on every project that pulls in a library new to that agent.
What Astryx does differently
Run `npx astryx manifest --json` and you get a structured JSON description of every CLI command, argument, flag, and response shape. Not prose an agent has to parse and guess at, a literal contract, the same idea as an OpenAPI spec for a REST API, applied here to a design system's own tooling. Pair that with the MCP server Astryx ships, and an agent scaffolding a screen calls a documented interface instead of pattern-matching on markdown it was never designed to be parsed by a machine.
The honest tradeoff
Astryx is in public beta, and the manifest plus MCP server are the newest, least-proven part of the whole system, which happens to be the exact part that is the entire selling point here. Eight years running inside Meta's own monorepo is not the same as eight years of outside teams pointing their own agents at that manifest against messy, non-Meta codebases. There is no independent track record yet of how well it actually holds up outside the environment it was built in.
Next step: try Astryx or the GitHub repo on a real screen before deciding it belongs in a stack. If you are evaluating dev tools for an AI-agent-heavy workflow and want a second opinion, write to us at hello@gattyworks.com.