Figma's Config 2026 pitch is that agentic tools should extend a designer's judgment. Dylan Field said it directly: "AI has lowered the floor, but it has not raised the ceiling." Designers raise the ceiling. The framing is correct. The architecture contradicts it.
The agent extension test asks one question. Can you describe how you think clearly enough that an agent applies it to a new case and you endorse the result? If yes, your thinking is a method. If no, you are delegating to something that cannot represent you.
Figma's answer is Skills. A skill is a set of instructions, written in plain English, that you teach the agent. You package your design philosophy and your conventions. The agent executes them. Your judgment, extended.
Extended by what? Figma AI runs a multi-model strategy. Different tools use whichever model fits best. Figma Make lets users choose between Claude, Gemini, and GPT. The model that interprets your skill and decides what it means on a new case belongs to Anthropic, Google, or OpenAI. Not to Figma. Not to you.
This would be abstract if Anthropic were a neutral infrastructure provider. It is not. On April 14, 2026, Mike Krieger resigned from Figma's board via SEC filing. Three days later, Anthropic released Claude Design, a standalone AI workspace aimed at replacing the need for designers in zero-to-one product creation.
Two remaining Figma board members are material investors in Anthropic. Findell Capital called for an independent investigation into whether Anthropic benefited from improper use of Figma's confidential information. The extension test breaks here. You wrote the skill. Figma hosted it. The model that interprets it is built by a company now competing with both of you.
Think about what a skill actually is. Figma calls it "a design philosophy, a signature aesthetic, a feedback lens, turned into something any teammate can implement, learn from, and build on." That is a method made legible. The value proposition is that your judgment becomes portable. But portable into a model you do not control, operated by a company that shipped a competing product three days after its CPO left your tool's board.
The margin pressure shows the dependency is structural. Figma's gross margin fell from about 92 to 86 percent over 2025 due to AI inference costs. The intelligence driving Figma's growth eats into its margins and comes from providers building competing products. Figma is paying its competitor to run the judgment layer of its own product.
A user on Figma's community forum put the frustration plainly: the agent is "constantly in the way" and "not something I think the majority of users want or need." The complaint is that the tool feels imposed rather than invited.
The gap between legible systems and judgment remains open. Design systems are now readable by agents. Brand context is not. 42% of designers cite lack of product and brand context as a top challenge when using AI tools for design work.
Skills are supposed to close this gap. They cannot close it when the model interpreting your method treats it as a prompt to be weighed against its own training, not a constraint to be followed.
Field described code as material for design, not its opposite. He is right. But material you shape through someone else's inference engine, while that someone builds a competing product three days after leaving your board, is material you do not control.
Figma's Skills feature is the best articulation of judgment-as-method I have seen from a design tool company. That makes the structural problem worse, not better. The clearer your method becomes, the more valuable it is to the model provider who interprets it. You are writing your playbook in a language only your competitor's infrastructure can read. So who owns the judgment layer that reads your Skills?