A few thoughts on the new Figma AI Agent after spending the last few days testing it.
A few thoughts on the new Figma AI Agent after spending the last few days testing it.
Short version: it's fun, occasionally useful, and already capable of accelerating certain workflows. But today it feels more like hiring a junior designer than a senior one.
The biggest limitation isn't the model quality itself. It's the surrounding system.
First, the agent doesn't appear to retain persistent project knowledge between sessions. Instructions such as "never overlap screens" or other project-specific rules often need to be repeated. Once a new session starts, much of that context seems to disappear. Whether that's an intentional design decision or a technical limitation, it creates friction for long-running design work.
Second, there is very little transparency around model routing. Figma is likely leveraging multiple foundation models behind the scenes and routing requests depending on task type, latency, cost, or other factors. That's reasonable.
But for professional users, I would rather have control.
If a task requires deep reasoning over large design systems, PRDs, specifications, and product context, let me explicitly choose the best model and charge me accordingly through Figma credits.
Enterprise users optimize for outcomes, not token savings.
Third, the system appears to operate within some form of scope or blast-radius constraints, but those controls feel underdeveloped.
Three things I'd love to see:
• A visible blast-radius estimate before execution. Show me what the agent expects to modify before it starts.
• A lightweight deterministic layer that detects ambiguous intent and asks clarifying questions before actions are executed.
• Most importantly: a dedicated knowledge layer.
Where is my design-system-rules.md?
Where is my figma.md?
Where is my Agent-Oriented Product Ontology?
Where do my PRDs, specifications, business rules, entity relationships, workflows, permissions, and UX constraints live?
Current AI tools seem heavily optimized around generating interfaces.
The next generation will be optimized around understanding products.
A mature enterprise application isn't just a collection of screens. It's a network of entities, relationships, permissions, workflows, business rules, and design constraints.
Without access to that knowledge layer, the agent is forced to infer too much. That's where hallucinations, inconsistency, and design-system drift start appearing.
For generating landing page variations, quick explorations, and lightweight A/B concepts, the current experience is already valuable.
For large enterprise products with complex domains, I think we're still early.
David Kossnick Dylan Field Kris Rasmussen Nadia Singer Figma