Figma Specs and Code Layers: What a Design Rulebook for AI Agents Means for Your Business

Figma used its first Config event in India, held on 14 October, to answer a question product teams have been circling all year: if AI agents are going to design and build software alongside people, who tells them the rules? The company's answer is Figma Specs, a Markdown document that captures how your product should look and behave, written once and read by designers, developers and AI agents alike.

The same event moved Code Layers into public beta, letting teams put working code directly onto the design canvas, and expanded Figma's MCP support with more than 20 new tools. Taken together, the releases show Figma repositioning itself from a design tool into the place where software teams decide what they build and ship, and where the agents doing much of the building come to check the rules before they act.

For business owners and managers this is more interesting than the usual feature news, because it goes straight at the biggest weakness of AI-assisted development: output you cannot predict. Let's sink in!

What Figma Announced at Config India

The headline release is Specs, a new Markdown-based editor for documenting a company's design language. A spec can live as a standalone Figma file or sit directly on the canvas next to the work it describes. Inside it, teams can link canvas elements and embed FigJam diagrams, code layers and finished designs, so the document stays connected to the actual product rather than drifting out of date in a wiki nobody opens.

Around Specs, Figma is building what it calls design system context. Teams can gather design libraries, code libraries and written guidelines into one centralized system, and AI agents draw on that system before they generate anything. Figma is careful with its wording here: the goal is more predictable output, not guaranteed obedience. AI remains non-deterministic, and the company does not pretend otherwise.

Beyond that, the canvas agent reached a wider group of users in the week before the event, Code Layers went from private to public beta, and Figma's Model Context Protocol support grew by more than 20 tools, including Open Code, Dia, Grok Bot and the Indian AI startup Sarvam. Payload, the open-source headless CMS Figma acquired in 2025, is being wired into Figma Make for building dynamic websites with real content management behind them.

How Figma Specs Actually Works

The idea behind Specs is almost old-fashioned: write things down. Most companies already have design rules, but they are scattered across brand PDFs, Slack threads, component descriptions and the heads of two or three senior people. That worked, more or less, when every change went through a human. It stops working when an AI agent generates a screen in forty seconds and has never read the Slack thread.

A spec collects those rules in one Markdown document: which components to use and when, how spacing works, what the voice of the product sounds like, what accessibility standards apply. Because it is plain Markdown, it is equally readable for a junior designer and for a coding agent connected over MCP. The same document that onboards a new team member now also constrains the machine.

Markdown may seem like a small detail, but it is doing real work here. It is the format AI tools already understand best, it versions cleanly, and it does not require anyone to learn a new editor. Figma calling it a medium rather than a file format is marketing, sure, but the practical choice is sound.

Code Layers Bring Prototypes Closer to the Real Product

Code Layers, the second big piece, lets teams add real code to designs on the canvas. You can ask the agent to convert a design into working code, then edit just a part of that code to iterate on the prototype. The feature has been in private beta since June 2026 and is now open to everyone in public beta.

The business value sits in prototyping. A static mockup answers what a product looks like; a coded prototype answers how it feels when you actually click it. Until now that second question usually required a developer to build a throwaway version. With Code Layers the throwaway version appears on the canvas in minutes, gets clicked, criticised and corrected, and the learnings flow into the real build.

It is still a beta, and we would not recommend treating code layers as production code. But as a validation tool for client work, it shortens the distance between an idea and an honest answer about whether the idea works.

The 17x Problem: More Code, Not More Shipping

The most quoted number from the event came from Figma's chief design officer, Loredana Crisan: there are seventeen times more lines of code being produced today than in the past, yet only about 1.3 times more releases. Her point was that most AI-generated code is written to explore what a product should be, not to ship it.

The figures are Figma's own framing rather than an audited industry statistic, so take them with some salt. The direction, though, matches what we see in real projects. AI tools make it very cheap to produce variants, and very easy to produce variants that quietly ignore your brand, your accessibility requirements or your component library. The expensive part has shifted from writing code to deciding which code deserves to exist.

That is the actual problem Specs is aimed at. Rules in a shared document will not make an agent deterministic, but they move quality control from after generation, where it costs review time, to before generation, where it costs a paragraph of Markdown. Industry context points the same way: OpenAI opened its Codex harness to all developers in September, and every vendor is giving agents more autonomy. Rulebooks are the counterweight.

MCP Tools and the Payload Deal Start to Pay Off

The 20-plus new MCP integrations deserve more attention than they got. MCP is the open protocol that lets outside AI tools read context from Figma, so a coding agent in your development environment can consult your design system instead of guessing from a screenshot. With Open Code, Dia, Grok Bot, Sarvam and others connecting, the spec you write in Figma starts following your project into every tool that touches it.

The Payload integration points in a similar direction. Figma bought the open-source headless CMS in 2025, and its downloads have reportedly grown sixfold over the last twelve months. Connected to Figma Make, it means a generated website is not just a pretty page but a site with structured, manageable content behind it. For small businesses this quietly removes one of the main objections to AI-built sites: that they fall apart the first time somebody needs to edit the content.

The venue mattered too. This was Figma's first Config in India, which the company says is its largest market by users outside the United States, working with 40% of the top 100 companies listed on the Bombay Stock Exchange. The centre of gravity of software production keeps spreading, and the tooling is following it.

Practical Takeaways for Businesses Building Digital Products

First, write your rules down, even if you are a ten-person company. A one-page spec covering your colours, components, tone and accessibility floor costs an afternoon and pays off every time a human or an agent touches the product. When Specs becomes generally available, that page drops straight in.

Second, if your team already uses AI coding assistants, connect them to design context through MCP rather than pasting screenshots into prompts. Context the agent can query beats context someone remembered to copy. Third, budget for the gap between generated and shipped. The 17x number is a warning: AI will flood you with plausible prototypes, and someone still has to test, harden and maintain whatever goes live. Treat generated code as a draft by default.

And if you outsource development, ask your partner a simple question: how do you keep AI output consistent with our brand and our standards? An agency that answers with a process, shared specs, design system context and human review is using these tools properly. An agency that cannot answer is shipping you whatever the model felt like that day.

Final notes

Figma gave no pricing and no general-availability dates for Specs or the Payload connection, and Code Layers remains a beta, so some of this is direction rather than product you can rely on tomorrow.

The direction is still worth acting on. The industry spent two years making AI generate more; 2026 is turning out to be the year of making it generate correctly. Specs, design system context and MCP integrations all serve the same thought: agents do their best work inside clear rules. Companies that write those rules down now will get consistent results from the same tools that give everyone else a lottery. That is not a bad trade for a Markdown file.