Figma Agent only works if your design system is written down for machines

The handoff is changing shape. For fifteen years, the design file was the thing we reviewed, signed off, and shipped against. Now Figma is pushing an Agent that reads your design system, and a growing group of designers are arguing the code itself should be the source of truth. If you run a team, this touches your hiring, your reviews, and what you call "done." Let me catch you up.

The deep cut

• The file is a drawing, not the building. The design file shows one state; the browser ships every other one, as the UX Collective piece argues.
• AI reads your design system, so write it down. Figma Agent needs a machine-readable contract, and John Rodrigues says 90% of systems lack one.
• Passing an automated checker is not passing. AI navigation code clears scanners while locking out keyboard and screen reader users, per William Hudson.

What the file never showed you

A design file shows one state, at one width, in one language, with the data the designer typed in. The product is everything else. The name longer than the box, the empty list, the four-second load, the person tabbing through with a keyboard. The UX Collective essay counts six states the file never drew, and notes that each gap gets decided by whoever writes the code, after design is called done, usually without a conversation.

So the file and the product drift apart. And the file keeps showing up in reviews as if it were the product. That drift is the real cost of a picture-first workflow, and AI only speeds it up.

Code as the material, not a second job

The old pitch, "designers should code," asked designers to become a slow second engineer. Many tried and quit. The argument now is different because AI writes the code. You describe the screen, its states, and its numbers, and the model builds it. You correct by looking at the result.

The author of the essay built 131 button variations in code and found that a vague adjective like "tactile" became a handful of real numbers an engineer could actually build. The point is understanding the material, not holding the chisel. Learn what wraps, what stretches, which states exist, and what a screen reader announces. None of that needs syntax. Jack Anglesea's team already swapped Figma handoffs for clickable HTML prototypes in one GitHub repo, trading some polish for fast domain feedback.

Why Figma wants your system written down

Figma Agent lives next to the canvas and handles migrations, token management, and audits. But the pitch underneath it is the hard part. John Rodrigues argues that around 90% of existing design systems were built for humans, not agents, so AI produces disconnected "slop" when it reads them. Guidelines split across Figma, Storybook, Slack, docs, and code cause the model to drift.

The fix is a machine-readable contract that tells agents exactly how to use components, tokens, and rules. Figma's new "specs" aim to be that shared layer between design and code. One example he cites hit 95% retrieval accuracy once a decision layer sat on top. The takeaway for you: an agent is only as good as the system you feed it, and your system probably is not ready.

The bug that passes the test

Here is where speed turns into a trap. William Hudson warns that AI-generated interface code keeps reproducing the same failures: mouse-only event handlers, a clickable div where a real button belonged, missing keyboard support, confusing focus order. The code looks correct. It can even pass an automated checker. And it stays unusable with a keyboard or screen reader.

His rule is blunt. Never ship AI-generated navigation without testing it with a keyboard and assistive technology. The tool speeds the work; it does not vouch for the result. Build on a plain HTML baseline first, then layer dropdowns on top. If you are moving reviews into code, this is the test you add, not the one you drop.

Three questions for your team

• Is our design system a machine-readable contract yet, or is the truth still scattered across Figma, Slack, and docs? If it is scattered, an agent will produce slop, no matter how good the tool is.
• What artifact do we actually review and sign off on, the file or the running product? Pick one and make it the source of truth, then regenerate the other from it.
• Who tests AI-generated navigation with a keyboard and a screen reader before it ships, and is that a named step in our process or an accident?