Anyone can build an agent. The harder problem...

# The Hard Problem of Building Agents: Figma’s Integration into the Canvas

The conversation around artificial intelligence in design has evolved rapidly, moving from basic generation to complex workflow integration. According to Yuhki Yamashita, building an AI agent is no longer the difficult part. The true challenge lies in making these agents work seamlessly with a team, rather than existing off in a siloed chat window. This philosophy drives the launch of Figma's new design agent, which is built directly into the multiplayer canvas where designers already think and work.


By embedding the agent directly in the canvas, Figma ensures that the AI shares the same context as everyone else on the team. This allows designers to explore more directions, automate tedious busywork, and gather feedback without ever leaving their file or losing the thread of their creative process. The agent is touted as being fluent in Figma, deeply understanding product design concepts, patterns, and Figma-specific use cases.

Furthermore, it takes full advantage of the multiplayer environment. Designers can run multiple agents in parallel or jump in to edit hands-on when it makes more sense. Because it is connected to a team's design system, the agent understands context, enabling it to handle component setup, fix layouts, and execute bulk edits efficiently.


### Community Reactions: Excitement, Skepticism, and Existential Reflections

The introduction of such a deeply integrated agent has sparked a wide range of reactions from the design community, highlighting both the immense potential and the complex challenges ahead.

**Context is Key**
Many designers see the in-canvas integration as a critical breakthrough. Rahul Sahu points out that context switching between a chat interface and the actual design file is where most AI-design workflows fall apart. By keeping the agent in the canvas, spatial context, layer relationships, and accumulated decisions are preserved. The multiplayer aspect shifts the mental model from merely "delegating to AI" to genuine co-creation.

**The Desire for Deeper Understanding**
While a design-aware agent is powerful, some are already looking toward the next evolution. Mateusz Cieply suggests that the next big step is an agent that is "product-aware"—one that understands component lineage, semantic meaning, and the data shapes behind the UI. Meanwhile, Uijun Park raises a practical question: given that companies spend years building design system infrastructure inside Figma, how well can the agent truly understand and consistently apply those specific guidelines? Prince Ngwenya echoes this sentiment, wondering if the agent can navigate a messy design file and make sense of a "junkyard" in relation to business goals, or if it is merely a brainstorming tool.

**Accessibility and Quality Concerns**
The potential for rapid automation brings specific risks. Viktoria Filippova raises a vital question for accessibility teams: when an agent performs bulk edits, does it preserve contrast ratios, focus order, and semantic structure, or does it simply provide a faster way to ship inaccessible work?

Others worry about the broader impact on the craft itself. Micah Tinklepaugh expresses concern that adding agents might actually slow design velocity and degrade quality, arguing that AI cannot truly automate the non-deterministic and dynamic problem-solving required in design. He warns that such tools might lead to designers becoming less talented over time.

**The Changing Nature of Craft**
Beyond the technical and practical implications, the shift is evoking a sense of loss for the traditional design process. Joshua Ruha captures a widespread sentiment: as work becomes increasingly easy to produce, the joy of struggling through ideas and crafting something by hand is diminishing. He reflects on the era before AI, noting that the process of discovery and experimentation gave the work its soul. The industry is standing at a crossroads, balancing the accessibility and speed of infinite variations against the raw, human aspects of traditional design.


### The Path Forward

Practical considerations also remain, such as token usage and costs, as questioned by Michael Tinglin. Furthermore, Morteza Ajidanpour notes that the utility of the agent may depend heavily on the underlying model, expressing a strong preference for the integration of capable third-party models like GPT 5.5 or Claude Opus over a purely native default.

As Figma rolls out its design agent, it is clear that the tool will change how teams operate. Whether it becomes a true co-creator or a source of friction will depend on how well it navigates the complex realities of design systems, accessibility standards, and the deeply human nature of creative work.