Figma's agent fills the components, but it still cannot tell you if the flow works

AI stopped being a side experiment in your tools. It is now shipping into the core of Figma, Canva, and Cursor, and it is aimed straight at the handoffs your team does every week. The question is no longer whether to use it. It is which parts of the job you let a machine run and which parts you keep in human hands. Let me catch you up.

The deep cut

• Automate the handoff, not the judgment. Figma's agent fills in components; a designer still decides if the flow works.
• An agent is only as good as the context you feed it. Dell's code layers worked because they pulled from the same React components developers ship.
• A model trained on design beats one trained on photos. Canva Craft plans layout and text before it renders a single pixel.

Context is the whole game now

The biggest jump in output quality does not come from a better prompt. It comes from what you hand the tool before you ask for anything. Evangeline's six-step workflow is blunt about it: feed the AI the real system, production code, PRDs, research notes, before you ask for a single screen. She pulls context in a fixed order and saves images for last because text costs fewer tokens.

Figma built this straight into the product. Wayne Lin's guide shows the agent pulling from your published library, so it uses your approved Card component and your spacing tokens instead of inventing new ones. The catch is honest in the fine print: guidelines work best when the agent builds from scratch, and Figma is still improving how it applies them to existing designs.

Code on the canvas moves the handoff earlier

Code layers let teams build working prototypes without leaving design. The payoff is real when the inputs are real. At Dell, Miguel Wermuth's code layers are built from the same React components developers ship, so a stakeholder can type into a text field and see the product respond. Merkle reports 30% faster iteration and up to 50% less time writing interaction specs.

The speed story gets loud fast. At Cox Automotive, a product manager who says "I'm not a designer" ran 19 versions with about an hour of hands-on time, roughly 90% time saved. Useful. But speed without judgment just means shipping the wrong flow faster. Derek Onay at Rocket Money used code layers to settle interaction and accessibility details early, which is the point. The tool moves the decision forward. A person still makes it.

A model that plans before it draws

Canva Craft is Canva's first in-house image model, and the interesting part is the architecture, not the horse race. It runs in two stages. First it plans the layout, spatial relationships, scale, hierarchy, and where text goes. Then it renders. That is why it handles legible text and precise object placement better than models that skip straight to pixels.

The training choice matters for your team. Canva trained Craft on design, human-made templates like posters, social posts, and menus, not photographs. So output can convert into layered, editable designs instead of a flat picture you cannot touch. It ranks fourth on Design Arena's leaderboard, behind three OpenAI models, so this is a strong contender, not a winner. The lesson travels: a tool trained on your kind of work beats a general one.

The audit nobody had time to run

The cleanup jobs that used to eat days are now scriptable. One designer connected Cursor to Figma through a read-only token and walked a five-year-old file through seven prompts: find the style guide, extract the real styles, then flag every layer using a raw value or a non-approved style. The output was a clickable report with direct Figma links to each violation.

Figma is packaging the same idea as skills. A /design-system-audit skill checks screens against your libraries and flags detached instances, hardcoded values, and accessibility gaps. These are the repeatable, rule-based checks worth handing off. They are boring, consistent, and easy to verify, which is exactly where an agent earns its keep.

What stays in human hands

Every builder in these pieces kept the same line. The agent fills in components and renders screens. People decide if the flow solves the problem. A Muzli designer put it plainly: AI may give you ten beautiful layouts, but UX knowledge tells you which one works, and "sometimes it just means we can make the wrong decision faster."

At Lufthansa Systems, Pank Juvekar used Figma Make to prototype live cockpit maps with real traffic data, testing with pilots again and again before a line of SwiftUI was written. The tool got realistic prototypes in front of experts earlier. The domain knowledge, knowing which zoom level clutters a pilot's map, stayed with the team. That is the split to copy. Automate the setup and the checking. Keep the problem and the call.

Three questions for your team

• Which of our weekly handoffs are rule-based enough to hand to an agent, and which ones hide a real judgment call we should protect?
• Is our design system actually feedable? If the agent pulled from our library today, would it produce our product or generic work?
• Before we buy another AI tool, can we point to the context, our code, tokens, research, that would make its output ours instead of anyone's?