The design system becomes the AI’s brief.

What Figma showed in June

At its Config 2026 conference in San Francisco, Figma presented its tool as an “intelligent canvas”, a space shared between designers, developers and AI agents. Several announcements point that way. Agent skills turn repetitive tasks into reusable actions, shared across the organisation. Generative plugins are created in natural language. Code layers let people work with code directly on the canvas: clone a repository, convert screens into editable layers, send changes back to the code. Figma also presented Motion for animation, as well as a gradual integration of Weave, which brings together some twenty generative AI tools.

In parallel, Figma’s MCP server has grown. It lets coding agents such as Claude Code, Cursor or Codex access a file’s components, variables and styles. In its remote version, it also lets them create or edit content in Figma.

The annual AI & Design report published on this occasion offers an interesting counterpoint. 57% of respondents believe design has gained in importance with the adoption of AI. Among developers, this proportion rises from 44% in 2025 to 65% in 2026.

Why context matters more than the prompt

A language model that generates an interface without context produces a plausible interface, in the statistical sense: it looks like the average of what it has already seen. It has no reason to respect your grid, your palette, your components or your accessibility rules.

The Model Context Protocol (MCP) changes the game by giving agents structured access to that context. Rather than guessing, the agent queries the design system: which component to use, which variables, which naming conventions. The quality of the result then depends directly on the quality of what it finds. A system that is consistent, documented and named the same way in Figma and in code produces consistent generations. An approximate system produces approximations, faster.

That is why several of Figma’s announcements are less about generation itself than about organising context: centralised design rules, reusable skills, connections between files and code repositories.

Tokens, a common language between design, code and AI

Design tokens are the elementary decisions of a design system, stored as data: colours, typography, spacing, radii, animation durations. They guarantee that the same decision applies identically in a design tool, a style sheet and an application.

In October 2025, the W3C Design Tokens Community Group published the first stable version of its specification (2025.10), which defines a common exchange format, with support for themes and modern colour spaces. For organisations, the benefit is twofold: tokens become portable from one tool to another, and readable by agents.

Naming matters as much as format. A token named after its function (primary text colour, alert background) conveys an intent that AI can apply correctly. A token named after its value (blue 500) says nothing about use. This is also where accessibility is guaranteed: text-background pairs validated for contrast, written into the system, propagate to everything that is generated.

What AI does not decide for you

Generative tools can produce variants, speed up exploration and automate adaptations. They do not know what distinguishes one brand from another, or which priorities a journey must serve.

The most visible risk is uniformity. Interfaces generated from the same models, with the same default conventions, end up looking alike. Differentiation then lies in the rules passed on to the AI, and in the choices deliberately kept human: information hierarchy, tone, brand moments, trade-offs drawn from user research.

A second risk is discussed less. Connecting agents to design files means opening access to unpublished mock-ups, confidential projects, sometimes real data used as examples. This access must be managed like that of any third-party tool: limited scope, identified accounts, regular review.

Preparing your system

For an organisation that already has a design system, or is considering building one, a few workstreams make the whole usable by AI.

• Formalise tokens in a standard format, with semantic naming.

• Align the names of components and properties between Figma and code.

• Document usage as much as appearance: when to use a component, and when not to.

• Build accessibility requirements into the components themselves.

• Version the system and designate who is responsible for it.

This work was already useful for human teams. It becomes the condition for AI to produce something that looks like you.

Frequently asked questions

What is a design token?

An elementary design decision (colour, size, spacing) stored as data, so that it is applied identically in design tools and in code.

What is Figma’s MCP server for?

It gives AI agents structured access to the components, variables and styles of a Figma file, in order to generate code or screens that comply with the design system.

Will AI replace designers?

Figma’s AI & Design 2026 report rather suggests the opposite: a majority of respondents believe design is gaining in importance with AI. Adaptation tasks are being automated, structural and brand decisions remain human.