Problem
The simplistic narrative says design agents will replace handcrafted screens. That’s a shallow reading. The launch of Figma Design Agent points to something more structural: the canvas becomes an operational surface where the agent understands components, tokens, layers, libraries, and team context.
It’s not just “prompt to design”. It’s “agent on canvas”.
Thesis
The value of Figma’s agent isn’t in producing more alternatives. It’s in connecting generation, direct manipulation, design system, and collaboration in the same space.
When the canvas is legible to agents, Figma stops being just a design tool. It approaches a product operating system.
Framework
The agentic canvas has four advantages:
- Visual context: the agent works on real layers, not on an external description.
- Design system: tokens, components, and libraries become operational instructions.
- Collaboration: humans and agents work within the same file.
- Direct manipulation: the user can manually correct when the prompt becomes inefficient.
Mini-case: a product team needs to explore three onboarding flows. Before, a designer prepared variants, another reviewed consistency, and development translated afterwards. With an agent in Figma, the team can generate directions, apply components, compare states, adjust copy, and push learnings to code without losing the canvas as a common source.
Measurable signal: time from product problem to reviewable flow with states, components, and clear criteria.
Posture: the agent doesn’t replace the design system. It punishes it if poorly defined.
Breathing: when generation is easy, the quality of the system becomes more visible.
What changes for design systems
A design system can no longer be just a visual library. It has to be legible by agents:
- clear names
- coherent tokens
- components with intention
- covered states
- documented patterns
- non-ambiguous usage rules
If the system is messy, the agent amplifies the mess at industrial speed.
Common mistake
The anti-example is asking the agent to “make a pretty screen” and then judging if it got it right. That reproduces the old problem of output without criteria.
The best use is more precise: “explore three structures for this problem using these components, respecting these tokens, and showing empty, error, and success states”.
The prompt stops being magic. It becomes product direction.
Protocol (3 steps)
- Audit your design system for agents. Names, tokens, states, variants, and rules.
- Prompt from layers and components. Not from vague abstractions.
- Evaluate with product criteria. Clarity, flow, states, accessibility, and consistency.
| Layer | Question |
|---|---|
| Canvas | What visual context can the agent read |
| Design system | What rules can it apply without asking |
| Product | What problem does the screen solve |
| Handoff | What happens when passing from canvas to code |
| Criteria | What does the human decide |
Related
- Brand System as Code: from guideline to executable system
- Stitch, Pomelli and the brand as an agentic pipeline
- Creative Governance: creativity, output, and system
Sources consulted
Next step
If your design system isn’t ready for agents, you’ll generate more visual debt, not more speed. We can audit it in a diagnostic.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.