What changed in Figma for squads that already design with AI in the flow

The problem every product squad knows isn't generating pretty screens with AI: it's what happens afterward, when that screen has to become code, follow the design system, and pass engineering review without turning into rework. Figma's recent updates, published in the official release notes between August 14 and September 3, 2026, address exactly that seam. It's worth reading the changelog straight from the source instead of waiting for a third-party blog summary, because several changes alter process decisions, not just the interface.

This piece separates what's noise from what actually changes the day-to-day of those who work at the bridge between design and dev in Brazil.

Generative plugins now deliver code and publishing

The most relevant change for the design and engineering relationship is in the September 1 update about generative plugins and shaders. At Config, Figma's annual conference, it enabled creating custom plugins and shaders just by describing what you want to the agent. Now that material is no longer an artifact locked inside the file.

The key points:

• Code viewer: lets you view and download the code for any shader or generative plugin you own, right inside Figma. This means what used to be the "designer's magic effect" becomes something auditable by whoever implements it.
• Animated and interactive shaders: you can ask the agent for an effect or fill that reacts to mouse movement. Nice in a demo, but this is where a trade-off lives: a reactive effect in a prototype is easy to approve and expensive to implement. It's worth agreeing beforehand on what's presentation flourish and what's an actual requirement.
• Publishing to the organization: on Organization and Enterprise plans, you can publish private plugins and shaders internally, standardizing effects and tools. It's the way to avoid ten versions of the same effect scattered around.

The technical detail the front-end team should test before trusting it: according to the note, when copying a frame with shaders applied to an external agent, on implementation "the shaders will render correctly in the generated React code." The official note's certainty doesn't eliminate the variable of the bundle and the project's actual design tokens, which the source doesn't cover. The practical recommendation is to treat the exported code as a reviewable starting point, not a final deliverable.

Admin-managed MCP changes who controls AI access

Two notes deal with the Model Context Protocol (MCP), the protocol that connects Figma to external agents like Claude. On August 24, company-managed authorization reached general availability (GA).

In practice, admins on Enterprise and Organization plans can now centrally control the connection between Figma's MCP server and AI agents through the company's identity provider, Okta for example. The operational UX gain is direct: users no longer have to authenticate individually or click through repeated consent prompts.

Why this matters beyond convenience: when an AI agent reads and changes design files via MCP, the security question stops being hypothetical. Centralizing authorization in the identity provider is what allows a company with a compliance process to say yes to using AI in the workflow. Without this control, many teams get stuck between banning the tool or allowing it without governance. The note states that support for more identity providers and more agents is planned, meaning today's range is still limited.

In addition, the September 1 note mentions MCP updates that allow viewing and editing generative plugins and shaders by a third-party agent via Figma's MCP server.

Auto layout that speaks CSS's language

The August 21 update on responsive spacing is small in its announcement and big in the friction it removes. Auto layout's auto spacing gained two new options, Around and Evenly, alongside the default behavior, which was renamed to Between. The naming choice isn't cosmetic: they're exactly the values CSS already supports in justify-content.

| Option in Figma | Behavior | Mental equivalent in CSS | |---|---|---| | Between | Equal space between items, none at the edges (renamed default) | space-between | | Around | Equal space between items, half that space at each edge | space-around | | Evenly | All spaces identical, including the edges | space-evenly |

The gain is one of translation: when design and front-end use the same vocabulary, the "but in Figma it looked different" conversation disappears. The evidence that this reduces rework lies in the terminology alignment itself, which eliminates the step where the dev has to interpret the designer's intent.

Along the same lines, the August 14 note brought responsive text wrap with two options, Balance, which distributes text evenly across lines, and Pretty, which avoids leaving a single orphan word on the last line. The usual rule applies here: none of this replaces checking the behavior with real text, especially in Portuguese, which tends to run longer than English and breaks layouts designed for EN screens. A visually balanced result with lorem ipsum turns into an accessibility and truncation problem once real content comes in.

Agent panel in a separate window and vector editing

Two smaller changes, but ones that affect the workflow of those who use AI all day. On August 26, the desktop app (macOS and Windows) started allowing the agent's chat panel to open in its own window, which stays visible when switching between tools and tabs. For anyone constantly switching between prompt and canvas, that's less context switching.

On August 24, vector editing gained an eraser and a paint bucket with drag-to-fill. You can now erase stray strokes and recolor multiple regions at once by dragging. The Shift+E shortcut switches to the eraser in vector editing mode and in Draw, and the same shortcut still works for the Prototype tab in Design mode.

What to adjust in the squad's process

Translating the changelog into process decisions, without relying on third-party interpretation:

• Definition of done: if the team has adopted generative plugins, add to the definition of done that the exported code goes through a front-end review before becoming a task. The Code viewer makes this possible, so it stops being optional.
• AI governance: if the company has an Enterprise or Organization plan, MCP's managed authorization is the prerequisite for the security team to approve agent use. Without an eligible plan, access remains individual and without central control.
• Shared vocabulary: standardize the use of Between, Around, and Evenly in the design system's components, so intent reaches the dev already mapped to CSS.
• Accessibility first: text wrap Balance and Pretty and reactive shaders need to be validated with real content in Portuguese and with contrast and legibility checks. A visual effect doesn't enter the definition of done without this step.

The thread connecting all these notes is the same: Figma is making AI output less opaque, whether by opening up the code, centralizing access control, or aligning names with what the browser understands. It's up to each squad to decide where that control becomes process and where it's still just a release note promise.