Your Figma file just became an agent's workspace

Figma let agents onto the canvas this month. Not a plugin, not an export — the actual file.

I’ve been saying that designers are building their own tools now and the file is where taste lives. This update makes that literal. Your components, variables, and auto layout are now the context an agent reads before it draws a single frame.

What did Figma actually ship?

Two tools and a way to encode intent.

Figma's announcement that the canvas is now open to agents describes it plainly: use_figma lets agents operate on the canvas using your components and variables, and generate_figma_design turns live app HTML into editable Figma layers. Skills sit on top — small, shareable instruction sets that tell an agent which steps to take, what conventions to follow, and what good counts as for your team.

It runs through the Figma MCP server, so Claude Code, Cursor, and other agents hit the same surface with the same permissions. The beta is free for now, with usage-based pricing later. The important shift isn’t the price. It’s that the file stopped being a handoff artifact and became the agent’s working directory.

Before, AI generated a screen elsewhere and you tried to make it match the system after. Now it starts inside the system and has to justify any deviation.

Why does an agent on the canvas change the job?

It moves the review loop from eyes to structure.

When an agent works in code only, you catch drift by looking at pixels. When it works on the canvas with real components and variables, drift shows up as structure — wrong token, off-system spacing, ad-hoc radius. You can lint that before a human squints.

You’re not redrawing what the agent did. You’re curating what it’s allowed to do. I wrote that why curation is the job now — the studio keeping 29% of output isn’t bad prompting, it’s good editing. The canvas just moves that edit point earlier and gives it rules to run against.

How do you give an agent taste it can follow?

You write down what you’d otherwise keep in your head.

Models don’t carry your preferences between sessions. Skills do. A good Figma skill doesn’t say “make it clean.” It says “use 8px base, 24/40 padding, 10px radius for cards, 6px for controls, zinc on near-black, no shadow above 3%.” Concrete, checkable, boring — and that’s why it works.

If you’re starting from a reference you love, extract it first. Measure type, color, and spacing, then name the trade-offs. File that profile where the agent reads it — a taste section in CLAUDE.md, a Cursor rule, or a shared Figma skill your team and agent both read.

The bar isn’t cleverness. It’s repeatability. FastCompany's look at the human work of giving AI taste shows even frontier labs teach models what good looks like by collecting judgments from real practitioners. Your Figma file is the small version of that effort — a few decisions, written down, so agents borrow them without reinterpreting each time.

What's the trap when your canvas is now code?

Speed that feels like progress.

The demo loop is intoxicating — prompt, generate, pull into Figma, tweak, push to code, repeat. You can now produce 30 screens in a day that all use the system and still feel like nobody decided anything. Auto layout correctness is not taste. Token compliance is not point of view.

Hold two gates.

System gate first. Does the output use only your components and variables, or did it invent new ones? Kill the invention or promote it deliberately. Undecided variants are how drift accumulates.

Judgment gate second. Does the page have clear hierarchy, one job per screen, and copy that sounds human? That’s the check Anthropic's guide to evaluating agents keeps coming back to — you don’t grade the prompt, you grade the output against what you said mattered before you generated it.

Keep generation parallel and judgment singular. Let agents draft in branches, pull the best two into Figma, and review against a short rubric on a schedule. Generation wants breadth. The canvas wants a file that says no quickly.

Figma gave agents the canvas. You still decide what stays on it.