Figma Make gets an AI coding agent that edits your codebase and opens pull requests
Design-to-code has spent a decade promising a clean handoff and mostly delivering a pile of throwaway markup. Figma's latest move at least narrows the gap. On May 28, Figma opened a closed beta for new Figma Make capabilities: connect it to your local codebase, then prompt edits on specific elements, adjust properties through a Figma editing panel, or describe changes in plain language — and an AI coding agent makes the corresponding edits in your actual code. When the result looks right, it can commit the changes and open a pull request without you opening a terminal.
What the beta actually does
The headline shift is that Make is no longer just generating a prototype in a sandbox. It's pointed at a real repository. You can select an element on the live interface and ask for a change, tweak spacing or component properties through a panel, or hand the agent a chat instruction, and it edits the underlying source. The loop ends at a pull request — meaning the change re-enters your normal review process rather than living as a one-off export you'd have to reconcile by hand later.
During the beta, these features don't consume AI credits, and Figma has said it expects to announce credit pricing later. The catch worth noting: at launch it's limited to the beta desktop app on macOS, with other platforms described as coming. So this is an early, gated look, not a general release.
Why the "no terminal" part matters
The interesting design decision isn't the agent — agents that edit code are everywhere now. It's the framing of the whole flow around a pull request. That keeps a human in the loop at exactly the point where it counts. A designer can drive a change visually, and the engineer who owns the codebase still sees a reviewable diff before anything merges. It's a more honest model than "the AI shipped it" — closer to "the AI drafted it, you decide."
It also reflects a broader theme from Figma's spring releases: AI moving from a side panel you consult into an action layer that touches the artifact directly. The same release wave leaned heavily on agents, MCP-style skills, and writing directly to the canvas. The throughline is reducing the number of times work has to be manually re-translated from one representation to another.
The skeptic's read
The thing every team will want to watch is the quality of the diffs on a real, non-trivial codebase. A generated pull request is only a time-saver if reviewing it is faster than writing the change yourself — and on a mature project with established patterns, that's a high bar. The structured-handoff problem design tools have chased for years was never really about generating code; it was about generating code that fits the conventions a team already lives by.
Still, anchoring the workflow to a PR is the right instinct. If you build or maintain a frontend, it's worth requesting beta access and pointing it at a low-risk corner of a project to see how the diffs read before forming an opinion. The promise is real; the proof is in the review.