Figma AI Agents: An Honest Review After Testing Every Prompt

Last week, Figma AI Agents finally showed up in my sidebar. If you do not see them yet, do not panic. Figma is rolling them out gradually, so you might have the feature today, next week, or not for another month.

The pitch is a big one: an AI agent that lives inside your file, reads your design system, and builds UI for you on command. As someone who builds design systems for a living, that is exactly the kind of promise I wanted to stress test. So I did. I ran the agent through a full afternoon of prompts, starting from the basic starter ideas Figma suggests and working all the way up to a fully connected production design system, and watched what it actually produced.

This is the honest write up. What works, what does not, and whether it is worth your time in its current state.

What are Figma AI Agents?

Figma AI Agents are a new set of AI tools built directly into Figma. Instead of generating a static mockup in a side panel, the agent works inside your file the way a collaborator would. You give it a prompt, it builds frames on the canvas, and in theory it pulls from the variables, styles, and components you already have.

If you have read about my Squarespace to Claude Code rebuild or played with Google's Stitch, the concept will feel familiar. The difference is location. This agent does not ask you to leave Figma. It builds where you already work, which is genuinely the right place for it to live.

How I set up the test

I wanted a clean test, so I started in a blank file. No existing components, no finished designs, nothing for the agent to copy or lean on. The only thing in the file was a variable collection: a proper set of colour, spacing, radius, and typography tokens, the kind you would set up at the start of any real project.

That setup matters. The whole promise of an in-file agent is that it respects your system. A blank file with a clean token collection is the fairest way to find out whether it actually does, with nothing else for it to hide behind.

Round one: the starter prompts

Figma suggests a handful of starter prompts, and I always recommend trying those first. I learned this lesson working with Stitch: the suggested prompts are tuned by the team behind the tool, so they tend to produce cleaner results than something you improvise on your own.

So I clicked the obvious one: create a dashboard UI.

First impression? Not great. AI slop, if I am being honest. It was the kind of generic dashboard you have seen a thousand times, with none of my tokens applied. I get noticeably better results out of Claude design and Claude Code.

But here is a pattern I have seen across every one of these tools, Stitch included: the mobile output is almost always better than the desktop output. So before writing the agent off, I pushed it toward mobile.

Round two: mobile flows beat desktop screens

Next prompt: build me a mobile onboarding flow for a finance app.

The result was a real step up. Mid fidelity wireframes, a sensible structure, and screens that actually connected into a flow. When I dragged one of the frames out, it held up responsively, which was a nice surprise. As a starting point for iteration, it was not bad at all.

Then I flipped it: build me a desktop onboarding flow for a crypto app. I wanted to see how well the agent translates the same idea from mobile to desktop.

It confirmed exactly what I suspected. Desktop is where these tools fall down. The layouts were looser, the structure weaker, and the polish dropped off compared to the mobile version. If you take one thing away from this test, let it be this: Figma AI Agents are at their best on mobile and at their weakest on desktop.

The variables and styles problem

Here is the part that mattered most to me, and the part that disappointed me most.

In every one of those early generations, the agent did not use my variables or my styles. Not the colours, not the type. I had a full token collection sitting right there in the file, and it ignored all of it. For a feature whose headline promise is that it works with your system, that is a rough start.

So I got explicit. I ran: build me a dashboard using the variables and tokens inside this file.

The result looked like progress at first. The agent told me it had referenced my tokens and styles. But when I inspected the output, the truth was uglier. Every value on the right hand side was a loose, unconnected hex code. The agent had name dropped my tokens without actually binding to a single one of them. Click into the text and the type styles were not applied either.

That is close to useless for real work. If I have to rebuild every fill and every text style from scratch just to connect them to my system, the agent has not saved me anything. It has handed me a picture of a design that I now have to redraw.

What happens when you connect a real design system

At this point I wanted to know whether the problem was the agent or my setup. A loose variable collection is one thing. A published, subscribed design system is another. So I tried two more steps.

First, I published the variables and styles as a library and subscribed to them from a fresh file, the way you would consume a design system across projects. The agent now reported that it found the variables and some of the styles. Closer. But the output still only picked up some of them, not all, and I would still have had to go through and manually reapply tokens. Not worth the effort.

Then I connected a full design system, the kind you would actually ship with. Real components for buttons, inputs, and cards, all wired to variables and styles. I asked for a dashboard and a transactions flow.

This time it clicked. The agent finally used my variables, my styles, and my components. The dashboard pulled real button and input components. The flow was cleaner again, and the type styles were mostly applied, with the odd broken token here and there (Claude Code throws a few of those too, so I will not hold it against the agent).

The lesson is specific and useful: the agent only respects your system when the system is made of components. Loose variables and styles get name dropped and ignored. Components get used. If you want Figma AI Agents to design with your tokens, you have to give it built components to work from.

What Figma AI Agents are actually good at

Here is the whole test, summarised in one table.

Read that table top to bottom and the sweet spot is obvious: mobile flows built on top of a real, component-based design system. That is where the agent earns its place. Everything outside that (desktop, single screens, loose tokens) is where it struggles today.

Figma AI Agents vs Claude

The honest comparison, since people will ask. My current workflow for AI assisted UI is Claude design for the first pass, then Claude Code to build it, then pushing the result back into Figma using my design system and a few Figma skills. I wrote about that broader build philosophy in the Squarespace to Claude Code rebuild, and reviewed the wider landscape in 7 best AI design tools in 2026.

Compared to that workflow, Figma AI Agents are not there yet. The output quality is lower, the token binding is inconsistent, and it needs full flows to look good. What Figma has that Claude does not is location. The agent lives inside the file, which removes the round trip entirely. If Figma closes the quality gap, that home field advantage could matter a lot.

Who should use Figma AI Agents right now

It is worth turning on and experimenting with if you are:

• Sketching mobile flows and want a fast, throwaway first draft
• Already working inside a mature, component-based design system
• Comfortable treating the output as low fidelity scaffolding rather than finished work

It is probably not for you yet if you are:

• Designing desktop interfaces or polished single screens
• Hoping it will respect a loose set of variables and styles
• Expecting production ready output you can hand off as is

The honest verdict

So, is it good? It is promising. Is it where I want it to be? No, not yet.

Figma AI Agents are at their best generating mobile flows on top of a real design system, and at their weakest on desktop, on single screens, and in files that only have loose tokens. For low fidelity iteration in the right setup, it is genuinely useful. For finished design work, it is not there.

The most encouraging thing is the direction. An agent that lives inside your file and designs with your real components is the right idea. Right now the execution is early. I will keep testing it as it improves, because the version of this that truly works would change how a lot of us start a project.

If you are building a product and want a design system that your team and tools like this can both actually use, that is exactly the kind of work I do. You can book a free consultation and we can talk through it.

Frequently asked questions

Are Figma AI Agents any good?

They are promising but early. The best results come from mobile flows built on top of a real, component-based design system. Desktop layouts, single screens, and loose token setups are noticeably weaker today.

Do Figma AI Agents use your design system?

Only partially. In my testing, loose variables and styles were referenced by name but never actually applied, which left unconnected hex codes all over the output. The agent only bound to my tokens reliably once they were attached to real components in a published design system.

How do I get Figma AI Agents?

Figma is rolling the feature out gradually, so it may not be in your sidebar yet. If you do not see it, you are most likely just waiting on the rollout rather than missing a setting.

Are Figma AI Agents better than Claude for design?

Not yet, at least in my workflow. I still get better results from Claude design and Claude Code, then pushing into Figma. What Figma has going for it is that the agent works inside your file with no round trip. If the output quality catches up, that advantage will matter.

Should I use Figma AI Agents for client work?

Treat the output as low fidelity scaffolding, not finished design. It is useful for a fast first draft of a mobile flow, but desktop work and anything you intend to hand off still needs a designer's hands on it.

Keep Reading

• From Squarespace to Claude Code: How I Rebuilt My Portfolio and Saved $528 a Year: the AI assisted build workflow I compare Figma AI Agents against in this post.
• 7 Best AI Design Tools in 2026 (Ranked and Tested): the wider landscape of AI tools designers are actually reaching for this year.
• Top Design Handoff Tools in 2026: once the design is done, this is the build-side stack that keeps your work intact.

---

About the Author

Sanjay Tarani is the Head of Design at DoxAI, helping entrepreneurs and business owners build scalable, user-focused digital products. Sanjay has led design system initiatives behind 50+ successful projects and has been recognised with the Website Wizard award. Sanjay brings experience from high-growth startup environments, including learning within the Startmate ecosystem, and shares practical insights on design, product strategy, and building profitable apps. Connect with Sanjay on LinkedIn.