Rollout is taking weeks. I would not tell a Head of Design the feature is on until it shows up in their toolbar. The object that holds all three is still a preview.
Two earlier notes are still in the plan. The agent went generally available on 6 Oct, and it spends AI credits. Motion styles landed on 30 Sep. I read the set the way I used to read a change against Aurora: what is on, who can touch it, and what a system team should refuse to fake before Q1 planning.
What you can open this month
Code layers are running code on the canvas. Open beta, rolling out to Starter, Professional, Organization, Enterprise, and Education plans, excluding K-12. You need edit access. Four ways in: capture a live page with the Figma Chrome extension, turn selected frames into a code layer, clone a connected GitHub repo, or start from an empty layer and prompt.
On paid plans, creating and editing inside a code layer is free through 4 Nov 2026. A single prompt can do several jobs. Only the code-layer part of that prompt is free. The agent's other work in the same prompt still spends credits. After 4 Nov, code-layer tasks spend credits too. On Enterprise, code layers go off only when an admin turns AI features off for the org.
Specs are a markdown document on the canvas. Open beta, rolling out over a few weeks to Professional, Organization, Enterprise, Education, and Government. View access is enough to read one. You need edit access to write one. You can place it next to frames, start a spec from the file browser, or fold a FigJam critique into one. An agent can draft it from outside context, including a codebase, through MCP, and you can turn a spec into a code layer. Using the agent on a spec spends AI credits. Hover the receipt after a prompt and you can see the cost.
Guidelines are markdown on a published library: rules, best practices, and the things you do not want the agent to do. Open beta on Professional, Organization, and Enterprise. When someone prompts with that library on, those files go to the agent in the background. A prompt can use more AI credits when the library includes guidelines, because that extra context is part of the call.
Code sources attach a GitHub repo or an npm package to the library. Closed beta, a small group. Assume you are not in it until the control shows up. You add the source, publish the library once, and teammates enable that library and prompt. The import is already done. The agent is meant to build and edit components and code layers from that code.
The centralized design system is preview only, on Professional, Organization, and Enterprise. The preview bundles libraries, code, and guidelines in one place, with a map of what a change touches and diagnostic reports. You cannot open that map for the whole company yet.
MCP can create code layers and read or write specs. Figma added 20 partners. generate_image spends AI credits. Your seat and plan still set the limit. As of 5 Oct, Organization and Enterprise plans can connect more than one GitHub organization to the same Figma plan. Use it when the system lives in one org and the product code lives in another.
Payload, also from 15 Oct, scaffolds a backend inside Figma Make and code layers: a CMS, a database, media, APIs, auth, and an admin panel. Supabase stays available, and existing Supabase Make apps keep working. Treat Payload as a prototype with a backend. The system you govern is still the library, the guidelines, and the package.
The rest of this quarter
We are a few weeks into Q4. The date inside it that I would put on a calendar is 4 Nov. If code layers have reached your paid plan, use the free window on a repo you trust. Clone the one the app installs, or capture a page that already uses the system. A blank frame is how you get a second button.
Write the anti-patterns down, in the library if your plan has guidelines.
"Do not detach this." "Do not ship the raw hex." "This pattern is for the marketing site. The app uses the other one."
We measured the repo version of that idea. An instruction file, AGENTS.md, tied to a +26.1 point lift on open-design-system-bench, across 898 graded generations. Figma guidelines are a different file, in a different product. I would still write the same rules in both places. Two documents that disagree teach the agent two buttons.
Shorter files cost less. A novel no one maintains is an expensive way to confuse the agent.
If you are in the code-sources beta, connect the package the app installs. A demo kit will teach the agent the wrong API. If you are outside that beta, write the prop names into the guidelines file now. Guidelines can carry the list this month. Code sources can enforce it later.
Then put one decision on the canvas. I've watched handoff fail between a frame and a page nobody on the engineering side opens. The frame says what it looks like. The page says what it must do. The agent sees whichever one you attached. A spec that points at the frame, the constraint, and the code is the first artifact both sides can share. Keep it short enough that a person will still read it. The agent will. Your reviewer might not.
If motion is the hole in the system, use what shipped on 30 Sep. You can save an animation as a style, back easing and duration with variables, combine styles into one pattern, and publish them to a library. The Figma agent, and a coding agent through MCP or Skills, can apply those styles. Motion is in open beta, and you can export a supported animation as Lottie or dotLottie. Publish the curves production already uses. Leave the experiments in the file. "Ease out, 200ms" becomes four different curves by the time it hits code.
What I would take into Q1 planning
Q1 is when your stakeholders ask what you did with this. By January, code layers cost credits. The centralized design system is still a preview, so I would leave it off the Q1 roadmap until you can open the map. The work I would put on that roadmap is the files you can write before then.
- Write the anti-patterns into the library guidelines, and mirror them in the repo instruction file.
- Before 4 Nov, make one code layer from the repo you trust, if the beta has reached you.
- Put the real prop names in the guidelines file until a code source is on your plan.
- Move one live decision out of a side doc and onto a spec beside the frame.
- Publish the motion styles production already uses.
That list is the contract. Scale, govern, and ship still means a person owns it. Figma gave the contract a place to sit. In Q1, the teams that filled it in will prompt against their own system. Everyone else will spend the quarter cleaning up a second button.
The job, from where I sit
I led Aurora at Nord. Dozens of teams, millions of users. The painful weeks were rarely a missing component. A rule lived in Slack. A prop meant one thing in Figma and another in React. Someone hardcoded a hex because the variable was annoying to find. We paid the inconsistency tax in review, then again in production.
An agent on the canvas has to pick a source. When the library, the guidelines, and the package disagree, it picks one and the repo picks another. The preview puts those three in one object, with a map of what a change touches. I want that map. I would not schedule a rollout around it.
We ran open-design-system-bench on four systems, 898 graded generations. Nobody came out AI-native. Aurora's composite on that sample was 63.4. The move that changed scores was getting an instruction file in front of the agent. Figma guidelines are the canvas version of that move. Keep them short. The extra context can spend credits.
If your Q1 is another component redraw, the agent will draw a kit next to yours. If the rest of this quarter goes into the rules, the real package name, and one spec beside a frame, you will have something the agent can follow when code sources leave the small group. Adoption still beats output. Start with the library the agent is allowed to see, and with someone who can say no when a prompt asks for a gray button by Friday.
Design and engineering, on the same canvas
On the design side, the canvas holds more than the frame. Set a spec beside it. Fold a FigJam critique into that spec when the critique produced a decision. Attach guidelines to the published library so the next prompt inherits the rule you are tired of repeating in Slack. The agent reads the spec and the guidelines. Put the decision there.
On the engineering side, a code layer is a prototype your PM can click. The component that ships still has to match the package: same props, same tokens, same states. MCP can draft a spec from the repo and place a code layer on the canvas. That shortens the argument. The merge still happens in the repo, in review, against the package the app installs.
Until a code source is attached to your library, the agent is guessing your components. The prop list in the guidelines file gives that guess a document. If the system and the product sit in different GitHub organizations, connect both on an Organization or Enterprise plan before you ask the agent to use the system.
If you admin an Enterprise plan, the switch is blunt. Code layers go off only when AI features go off. Make that call this quarter, with the people who own the credit budget, before Q1 prompts spend on code layers by default.
Which of the three is up to date on your team: the library, the guidelines, or the package?
The library, as heat
Twelve Nexus components from this site, passed through the heatmap shader. Darker pixels run hotter. Filled controls burn. Outlines barely register.
I checked the Figma release notes and the Config India 2026 help article on 18 Oct 2026. Plans, betas, and the 4 Nov credit window come from those pages. If your admin console disagrees, trust the console.