Figma is turning its design tool into something that feels a lot more like a teammate than just a canvas, with a new AI-powered “design agent” that lives directly inside Figma Design and works alongside you on your files.

If you’ve been following Figma’s AI story over the past year, this launch feels like the missing piece. First, Figma opened the canvas to third-party agents via its MCP server, letting tools like Claude Code or other MCP-compatible clients read and write to your Figma files. Now, instead of talking to AI through a separate window or plugin, you get a native Figma agent that sits on the canvas and in the left rail, understands your components, tokens, and design system, and can actually move pixels for you.

The promise is pretty simple: you shouldn’t have to choose between speed and precision. Historically, AI design tools have been great at spitting out quick mockups but annoying to correct; they often ignore your design system, produce layouts that break your grid, or generate components that don’t exist in your library. Figma’s agent is explicitly trying to solve that by being “fluent in Figma,” as the company describes it, and by being deeply aware of your actual files instead of treating them as static images.

From a workflow perspective, Figma is positioning the agent as a true collaborator on the canvas rather than a one-shot “generate a UI” button. You can start a prompt from any design layer, ask Figma to explore variations, and then keep editing while the agent continues iterating in parallel. That means you can, for example, select a checkout flow frame, ask for three different stylistic directions — organic, modern, retro — and watch the agent generate alternative screens that still pull from your existing components and variables. Because everything it creates lands as normal Figma layers, you’re free to move, override, and refine just like any other design work.

Figma clearly expects the agent to play two big roles: exploration and busywork. On the exploration side, it’s built to help teams “go wide” fast, so instead of spending hours making small variants on a single idea, you can use prompts as a multiplier. Think of prompts like “Make a horizontally scrolling image carousel for mobile, with different title treatments above and below the image” — that’s exactly the kind of example Figma uses in its announcement, and the agent responds with multiple options wired to your system. You can also go deep on a single direction, asking the agent to refine an existing flow while staying aligned with your tokens and frequently used components.

On the busywork side, this is where the agent starts to feel like infrastructure for serious teams. Anyone who maintains a design system knows the pain of renaming variables across a file, swapping components across dozens of screens, or propagating a padding change through an entire flow. Figma’s agent is explicitly designed to handle those sorts of bulk, context-sensitive edits. You can chat with it to “set all chip components to their active state,” “find all text larger than size 10 and change the font to Inter,” or “convert these screens to dark mode without wrecking accessibility contrast,” and it will walk the file, apply consistent changes, and keep everything tied to your system.

The same applies to content. Instead of manually replacing lorem ipsum, the agent can populate frames with more realistic copy and imagery at scale. For teams that care about coherent prototypes, this matters: better content makes it easier to get meaningful feedback from stakeholders who don’t speak in design jargon but understand flows and copy. And for design-system owners, the agent can bulk-update descriptions, tags, and usage notes, essentially helping document your components and states as part of everyday work rather than a separate, painful project.

What makes this agent more interesting than yet another AI sidebar is how it hooks into the broader Figma and MCP ecosystem. Figma’s MCP server, which is in beta, is an implementation of the Model Context Protocol that lets external AI tools access semantic design context directly from your files. That’s what powers things like Claude Code to Figma, where you can send a live running UI from your browser into Figma as editable layers, iterate on the design, and then push changes back to code. The new native agent is designed to sit in the middle of that loop: you might start in Figma Design, use the agent to explore flows and states, send the resulting frames to Figma Make or an MCP client to generate code, and then bounce back to the canvas to refine the visuals.

In other words, Figma isn’t just adding AI to the canvas; it’s trying to connect code and design into one continuous, agent-augmented workflow. The company has been saying publicly that “the future of design is code and canvas,” and the combination of the MCP server with an on-canvas agent is how that vision starts to look real.

Another area Figma is explicitly targeting is feedback. Design teams already treat Figma as a shared workspace where comments, crit notes, and stakeholder reactions pile up, often across multiple frames and pages. The agent can read that context and help turn raw feedback into action. You can ask it to summarize a long comment thread, group feedback by theme, or even “create a new revision of this design that incorporates the feedback for this profile.” Figma also suggests using the agent to pressure-test designs from different perspectives, like asking how a revenue-focused VP might react to a particular flow. It’s an attempt to make feedback less of a static backlog and more of a promptable input into the design process.

From a user-experience standpoint, Figma is leaning hard on the idea of “no toggle tax.” Instead of jumping between your design file and a separate AI app or browser tab, you stay inside Figma the whole time. The agent is embedded where the work happens, and because Figma is already multiplayer, teammates can see what’s going on, riff on outputs, and keep context in a single file. This is a subtle but important difference from exporting screenshots into an AI tool or copying prompts into a chat window — the agent is meant to feel like an integrated part of the canvas, not a bolt-on.

On the access and pricing side, this is rolling out as a limited beta. According to Figma’s help center and release notes, the agent began rolling out on May 20, 2026, to eligible Figma Design users and will expand gradually over the coming weeks. During the beta, usage won’t consume AI credits; those will only apply once the feature reaches general availability. The agent will be available to Full seat users on Professional, Organization, and Enterprise plans, while Collab and Dev seats can use the agent in drafts. Starter, Education, and Government plans are not included. Figma is using a waitlist model for early access, and joining it doesn’t guarantee entry, but selected accounts will get an email when they’re in.

This launch also fits into a broader trend: design tools are moving away from single “AI features” and toward AI that is deeply wired into design systems and code. Figma’s MCP server gives AI agents real design-system context, so when they generate or edit designs, they’re pulling from your actual components instead of a generic UI library. For large teams with mature systems, that’s the difference between AI that produces pretty but unusable mockups, and AI that can truly accelerate shipping production-quality interfaces.

For designers, the practical question is whether this agent will change how they work day to day. If Figma delivers on what it’s promising, you can imagine a fairly natural evolution: use prompts to sketch flows and explore directions quickly, switch to hands-on editing for fine-grained craft, lean on the agent to handle bulk system updates and content filling, and bring in MCP-connected tools when you need to sync design and code. Crucially, Figma keeps emphasizing “direct manipulation” — the idea that editing in the canvas should remain faster and more natural than endlessly tweaking prompts. The agent is there to amplify that way of working, not replace it.

The bigger implication is that design teams may start to think less in terms of static files and more in terms of living systems connected to code, data, and AI context. With MCP, external agents can already read and write to Figma files; with the design agent, a native assistant can navigate your canvas, apply your system, and orchestrate changes inside the tool itself. For organizations that are serious about design quality and velocity, that combination could turn Figma from a place where work is documented into a place where work is constantly being transformed.