Figma launches Design Agent and Dylan Field explains why designers shouldn't outsource their thinking to AI

Figma launches Design Agent

Figma launched its Design Agent today, and Dylan Field frames the product in terms that will feel familiar to anyone who has watched AI tooling oversell replacement and underdeliver on usefulness: the goal is to empower designers, not displace them.

The agent ships with faster performance, higher-quality outputs, improved ability to pull context across files, and expanded agentic capability on the canvas. Field says design system handling has improved meaningfully, though he acknowledges more work remains there. The underlying approach combines Figma's own first-party models with frontier models, with the team focused on making the combination useful for how designers actually work rather than optimizing for demo quality.

“We've launched Figma Design Agent today... We're not trying to replace designers. We're trying to make it so they are maximally empowered in their work... I think that the more important thing you can do is to actually use the model to learn, to try things you wouldn't have tried before, to be more daring, and to kind of expand possibilities rather than contract them.”

The canvas as the interface

Field's broader argument is that text is not the final representation of creative work. Non-developers — designers, PMs — are not at home in the terminal or the IDE, and the canvas offers something those environments don't. His bet is that reasoning models become part of the creative process rather than the whole of it, with humans and agents working together rather than one replacing the other.

On model strategy, Field is explicit that Figma is not trying to build a general-purpose AI. The value comes from deep domain focus — years of understanding how designers work, what Figma's product needs to do, and how to teach models the Figma environment before asking them to do good design work. He also acknowledges the tension with the so-called "bitter lesson": you don't want to fight against general model improvements; you want to layer domain expertise on top of them.

On AI-generated slop and design quality

Field is candid that AI-generated output in design, as across most categories, still skews toward the generic. Figma is working to reduce that, but he makes a more pointed observation about taste: different people in different domains define quality differently, so the real challenge is understanding what a specific user is going for and delivering that, not optimizing for some averaged standard.

Outsourcing your thinking

The sharpest part of the conversation is Field's position on how people are using AI. He says he spent roughly three weeks rewriting an internal post on AI writing policy before publishing it, because he wanted to get it right. The policy he landed on: any Figma communication that contains AI-generated writing of any substance should be marked "written with AI" at the top. Spell-checking and light proofreading don't count. The goal is to preserve trust that when someone sends something, they actually thought about it.

Field's broader concern is that people are using models to avoid thinking rather than to think better. The more valuable use, in his view, is using AI to try things you wouldn't have tried before, to learn, and to expand what's possible — not to compress the effort of understanding a problem you haven't actually worked through. He draws a line that applies well before AI: if you can't explain something, you don't know it yet.

Field also introduces Weave as a Figma initiative aimed at helping users shape model outputs like clay — iterating toward their own vision rather than chasing a one-shot result that rarely lands where it should.

His closing note is less product and more philosophy: in a period when everything is becoming less mysterious, there's an opportunity to lean into novelty and strangeness rather than converge to the mean. Whether that translates into a product roadmap or stays as a design principle, he leaves open.