Design system fidelity now flows through AI agents
Published 1 September 2026 · Source dated 7 August 2026
IMG.LY's August roundup of AI design agents notes that Figma and Canva both now expose design generation through MCP — meaning agents can read and write into your actual design files, tokens included.
Per IMG.LY's comparison of AI design agents, Figma has opened its MCP server to external coding agents (Claude Code, Codex, Cursor), and Canva now exposes its own design generation through MCP inside ChatGPT, Claude and Copilot. That means agents aren't just generating mockups in isolation anymore — they can touch your actual component libraries and tokens.
**Label: adoption.** Multiple major platforms have independently shipped this capability, which is a stronger signal than a single vendor's roadmap slide.
The catch for design systems teams: an agent can only respect your system if it's legible to a machine. Messy naming, undocumented variants and orphaned components will get bulldozed or ignored, not respected.
Coach
Open your most-used Figma or Penpot component and check if its name and variants would make sense to someone (or something) that's never seen your file before.
Google's Material Design 3 popularised dynamic, token-based theming back in 2021, and it's still what most design-system setups get measured against. If you're building variables in Figma or tokens in Penpot, M3's structure is worth studying, not because it's new, but because it's proven.
open Material 3's token docs and compare their primitive-vs-role token split against your own Figma or Penpot library — note one gap you could fix this week.
Creative Market's 2026 trend forecast flags 'Glasscore' — blur, glow and translucent layers — as a rising visual style. For teams running token-based design systems, that means more theming flexibility, not just flat colour tokens.
Open your team's (or a starter) design system file and check whether blur/elevation tokens exist — if not, sketch one variant using Figma or Penpot's effect styles.
Some teams are wiring AI coding assistants into their design token workflow, using rules and MCP connections to keep components consistent. It's promising but still fragile in practice.
pick one component in your (or a practice) design system and check whether its tokens are named consistently enough that an AI tool could use them without guessing.
