Vendors Sold Fidelity This Week. Design Judgment Still Has No Benchmark.

Two launches this week share a premise: that an AI agent which matches a template is an agent that designs well. OpenAI positions GPT-6 Astra as a model that adheres to existing formats and produces “well laid out” slides OpenAI. Figma shipped generative plugins and shaders that render deterministically from agent-written code Figma Engineering. Both are real progress on perception. Neither touches the harder question: does the layout communicate? This roundup is based on official release notes, engineering blogs, and press coverage — we did not run these tools hands-on.

What agents can now perceive — GPT-6 Astra and template fidelity

OpenAI says GPT-6 Astra, launched September 3, is “our best model for adhering to existing templates and producing slides that are well laid out” OpenAI. That is a vendor claim, not a benchmark result. The demo shows a few template slides becoming decks “correctly formatted for your business standards” OpenAI. API pricing is $10 per million input tokens and $50 per million output tokens CNBC. A second demo takes a Blender house and produces a walkable Unreal Engine 5 scene OpenAI.

What an agent can now perceive is a token stream that maps to layout structure — headings, bullet levels, image placeholders. That is genuinely new. But template conformance is a syntactic match. The agent sees that a slide has a title and three bullets in the right slots. It does not see whether the title carries the argument or whether the third bullet is visual noise. OpenAI’s own framing — “adhering” and “well laid out” — treats format as the criterion. For an engineer building a design agent, the takeaway is to separate these: measure template-match and communication as distinct axes, because the model’s marketing conflates them.

What tools change the audit surface — Figma generative plugins and shaders

Figma’s September 1 engineering post describes an agent that builds generative plugins and animated shaders from prompts Figma Engineering. Shaders react to mouse movement. The release notes confirm that publishing works to Community and private orgs, and that third-party agents via the Figma MCP server can view and change them Figma Release Notes. Critically, shaders render correctly in the React code Figma generates Figma Release Notes. Rogie King’s engineering voice: “what ends up on the canvas actually looks like you,” where building means moving between design and code Figma Engineering.

This changes the audit surface. Agent output is no longer pixels that require visual inspection. It is code — shaders, plugin logic — that renders deterministically. A reviewer can diff the generated code, check for side effects, and trace how a mouse event drives a visual state. That is an improvement over eyeballing a screenshot. But code correctness is not design judgment. A shader can be bug-free and still communicate nothing. The practical implication: build your audit pipeline around the code artifact, not the rendered frame, and treat code quality as one signal among several.

What criteria become computable — opacity variables and the always-visible agent

Figma’s September 3 opacity update makes a design property machine-checkable. Opacity can now be updated at scale from the variables modal without detaching instances, and it can be set via a number variable scoped for color variables — applying to disabled states, overlays, and scrims Figma Release Notes. This turns opacity from a per-layer eyeball check into a token-graph edge. An agent can verify that every disabled state references the same variable, that no hardcoded 0.4 breaks the system. That is a hard, computable criterion.

Separately, the agent chat panel can now open in its own window on macOS and Windows desktop, movable and always visible Figma Release Notes. This is an agent-inside-the-loop posture rather than a modal interruption. For a builder, the takeaway is concrete: treat opacity as a graph edge, not a number; and design your agent’s presence so it can observe state continuously rather than acting in discrete bursts.

Who audits agent action — Enterprise-managed MCP authorization

On August 24, Figma made enterprise-managed MCP authorization generally available. Org and Enterprise admins centrally manage which AI agents — Claude, for example — may connect to Figma via MCP, enforced through the identity provider such as Okta Figma Release Notes. There are no per-user consent prompts; more IdPs and agent providers are planned Figma Release Notes. This is governance catching up to capability. An agent’s identity is now revocable at the SSO level, which means actions are attributable to a specific agent principal. For your own systems, log which agent identity touched which file — that is the audit trail that makes design-agent behavior reviewable at all.

Resolution

Fidelity is the easy 80 percent. Matching a template, referencing a variable, rendering a shader — these are computable, and vendors sold them hard this week. Judgment is the unquoted 20 percent: whether hierarchy serves the content, whether contrast survives real use, whether spacing reads as intention. No release note benchmarked that. Agents can now perceive tokens and variables; they still lack a criterion for communication. Build the axis-2 checks yourself, and treat any vendor claim of design skill as a claim about format until proven otherwise.

Rule: format-match-is-not-communication — a computable rule the weekly Empire Design Scan can check: Never score a layout as good because it matches a template. Check two axes separately: (1) format match (token/class/template conformance, computable) and (2) communication (hierarchy, contrast, spacing consistency against the content's intent). Treat a template-conformant but communication-broken layout as a FAIL, not a pass. Vendors sell axis 1; agents must build axis 2.