Performance Gaps in AI-Generated UI

Figma AI agents currently function best as a tool for low-fidelity iteration rather than high-fidelity production. The tool demonstrates a clear performance disparity between mobile and desktop outputs: mobile flows are consistently more coherent and usable, while desktop designs often result in "AI slop" that requires significant manual cleanup. Furthermore, the agent struggles with single-page generation; it performs significantly better when tasked with creating entire user flows, where it can better contextualize the layout and component relationships.

Design System Integration Challenges

The primary friction point for professional designers is the agent's inability to reliably respect existing design systems. In initial testing, even when variables and styles were present in the file, the agent frequently ignored them, defaulting to hardcoded hex values and detached text styles.

To achieve consistent application of a design system, the following conditions are required:

• Library Connection: The agent requires a published design system library to be explicitly subscribed to within the working file.
• Component Utilization: The agent appears to only apply variables and styles correctly if it is actively utilizing components from the connected library. Without components, the agent fails to map design tokens to the generated elements, forcing the designer to manually re-apply them.

Strategic Recommendation

While the technology is promising for rapid wireframing, it is not yet a replacement for manual design workflows. For high-quality results, the author suggests that using external models like Claude to generate code or design structures—and then manually implementing them into Figma using established design system expertise—remains a more reliable and efficient process than relying solely on native Figma AI agents in their current state.