Built from fetched primary and vendor sources; no tools were run hands-on.
The revision round is the unit of agent learning
The sharpest sentence about design agents this week came from the company selling one. In an interview with the Seoul Economic Daily published October 31, Rene Lo (spelled “Renee Loh” in the piece’s photo captions), Figma’s VP of sales for Asia-Pacific, named the problem plainly: “The problem is quality. Designing something at a quality level you can actually ship now takes far more time and effort.”
The numbers Figma put alongside that admission are vendor-supplied — a Figma survey, as reported by the Seoul Economic Daily, with no methodology, sample size, field dates or question wording published. Among them: 51% of South Korean developers used AI on every project over the past year; 27% of respondents affected by productivity changes said time to launch grew longer; 47% said their AI tools lack design context; and 62% said they still need six to 10 rounds of revisions before AI output can be applied to a live service.
The mechanism claim survives the discount. If output needs six to 10 rounds before it ships, then the round — not the draft — is where the work happens, and a round is the atomic unit of agent learning: a human correction typed into a chat box, satisfied locally, then evaporated. Figma’s own engineering post, “How we’re training a Figma-fluent editing model” (October 29, just before this window), makes the same move by treating the undo rate as a quality metric. Everything else in this digest is an attempt to shorten, cheapen or make legible one round.
What agents can perceive: the canvas became executable
On November 2, Figma’s release notes reported that shaders now export “in up to 4096px max resolution” and can be carried into code layers created through the “build this with code” path. A shader renders inside a code layer when that layer is set to “Play in the top-right corner of the layer”. The release states: “Shaders are now generally available to all users.” The gate is stated just as plainly: shaders in this form are “Only available in the browser via Chrome version 157, rolling out this week” — a rollout state, not a ship date.
The code layer itself is the perception upgrade. Figma’s help documentation defines it as “an interactive prototype backed by running code that sits directly on the canvas” (Figma help, code layers), and describes several creation paths; the feature is “gradually rolling out in open beta”. An agent that can emit a running artifact, rather than a picture of one, has something it can execute and compare against — the precondition for any automatic check.
A vocabulary agents can check: Weave’s glossary went canonical
A concept an agent cannot name is a concept it cannot be graded on. On October 31 the Figma blog’s Weave explainer fixed four definitions verbatim: a Node is “An individual building block of a workflow on the Weave canvas”; a Workflow is “Connected nodes that turn your starting material into a finished result”; a Branch is “A separate path in a workflow for exploring a different direction from the same starting point”; and a Weave tool is “A workflow packaged so someone can run it without working with the underlying nodes”.
Two sentences in that post matter more than the glossary. “You can also run the same prompt through different AI models and compare the results side by side” turns model choice into a comparable variable rather than a vibe. And “Each version stays connected to the steps that produced it” is a provenance claim: the artifact carries its own lineage, so a reviewer can ask why an output looks that way without reconstructing the session from memory.
The publishing side landed in the window too: on October 30, Figma’s release notes added internal publishing of Weave tools, so published tools appear in the “Tools tab in Figma alongside Weave tools built by Figma and the Figma Community” (Figma release notes; see also Figma help, Weave tools). The caveat is a plan gate: “Internal publishing is only available on Organization and Enterprise plans.”
Context on tap: connectors widened perception — with a 16 MB ceiling
On October 30, Figma made Google Drive connectable to the Figma agent and to Make, with the release notes naming the intended uses as to “summarize research insights, reference real data, or pull in assets to edit on the canvas”. The adoption line: “Connectors are now available on all Figma plans, including Starter.” They are added from “the plus menu in the prompt box.”
The mechanics are documented in Figma’s help page on verified partner MCP connectors: the agent acts as an MCP client calling tools exposed by an MCP server, and “the agent decides which connector and tools to use”. Consent is explicit — a prompt names the connector and the tool it intends to call. Two limits bite in practice: downloads under an “Always run” setting may fail above 16 MB, with images over roughly 5–6 MB likely to exceed that; and “Each person needs to add and authenticate the connectors”, so context access is per-seat configuration rather than a workspace switch.
What criteria agents can compute: three gates you can run this week
Design Systems News, Issue 222 (fetched November 2, 2026) supplied most of the raw material: token linting “should occur at multiple stages of the workflow, from the design tool to pre-commit hooks and the CI pipeline”; component specs “often lose, distort or invent design intent during transformation”, which the newsletter proposes to test by “beaming specs back into Figma and comparing results” — a round trip; and, crucially, “perfect losslessness isn’t the goal; naming the gaps enables teams to build stronger contracts”. On the token side, the DTCG dimension type carries “explicit value and unit pairs”, aliases keep scales synchronised, and the issue’s operating advice is a design constraint worth adopting: “A scale you can explain in one sentence is one your team will actually follow.”
Three gates, each falsifiable, none requiring a purchase.
| Gate | Input | Checkpoints | Fail condition | Source |
|---|---|---|---|---|
| Token lint | Token source of truth, including aliases and dimension tokens | Three: design tool, pre-commit hook, CI pipeline | Any unresolved alias, or any dimension token without an explicit unit, anywhere | Design Systems News, Issue 222 |
| Round-trip spec diff | Component specs exported and re-imported into Figma | Each spec change | Invented properties ≠ 0; gaps not classified as lost, distorted or invented | Design Systems News, Issue 222 |
| Registry admission | Agent-emitted interface payloads | Every render | A payload referencing a component id not in the registered set — hard error naming the id | PPC Land |
Issue 222 also carried a sponsored item, labelled as one: “DS ARC: The AI readiness checker - brought to you by zeroheight”. As described in the issue, it runs “a sample of your docs against a rubric based on the Design System Documentation Spec, then point an agent at the docs with eight real jobs”, showing transcripts. The rubric content and the eight jobs are not published, so treat the outputs as the sponsor’s claims rather than an independent benchmark.
What agents may assemble: Google’s A2UI keeps the registry in your hands
PPC Land reported on October 31 (material dated October 29) that Google published a joint guide with CopilotKit showing agents answering with “selectable lists, date pickers, maps and booking forms”. A2UI is described as “Google’s open standard for describing an interface as data”, carried “over a second open protocol called AG-UI”.
The portable part is the pattern: “An agent can only assemble components the application has registered in advance”, which PPC Land reports as a design Google presents as “the core of its security model”. That is a criteria layer expressed as an allow-list rather than a prompt instruction — the agent’s output space is bounded by what the host application has admitted, not by what the model was told to avoid. Invention is rejected at admission time, not repaired in revision six.
State the limits plainly. The guide “gives no adoption figures, names no production deployment”. On versioning: “A2UI v0.9.1 is the current production release; v1.0 is a release candidate”. Nothing here establishes who runs it.
How agents learn from feedback: undo rates, head-to-heads, and version logs
Learning from feedback requires feedback that outlives the session, and three items this fortnight show instrumentation arriving from different directions.
The first is pre-window context, dated October 29. Figma’s editing-model post reports that typical tasks are “about 20% faster” between beta and general availability; that the new agent “beats or matches the previous production agent 62% of the time” in head-to-head testing with expert designers; and that “User ratings are 33% higher, and users undo the agent’s work 60% less often” (Figma blog). Not established: the identity of the previous production agent, the sample sizes, or what “user ratings” measure. Do not fill those in. The undo rate is the figure to keep, because it directly proxies a round not spent.
The second is version history, dated October 27 and therefore context: each new version of generative plugins and shaders receives an agent-generated description forming a change log, and the interface shows which version is currently published (Figma release notes). A change log written by the system is feedback that survives the chat window.
The third is access, dated October 26, also context: all Weave video and image models are callable through the Figma MCP — “models like Seedance, Kling, and Nano Banana” — and generation “will consume Weave credits” (Figma release notes). Cheap iteration is a feedback-rate question, and credit consumption is the meter on it.
What tools they need: the design-system layer got its own agents
Supernova shipped an agent for the design-system layer, and it is pre-window context: on October 26, Supernova announced Supernova Agent in public beta “on every plan”. It “reads, creates and updates your design system: documentation, tokens, components, Figma links, skills, contexts and export pipelines”. Writes are confirmation-gated — “Publishing and deleting wait for you: The agent proposes these and waits for your confirmation” — and an editor MCP exists. The cost line is operational, vendor-stated: “Each message uses 50 workspace credits from your workspace’s monthly allowance.”
The public-sector side moved on October 30, when the VA.gov Design System’s what’s-new page added a “design-staging-review skill for Collaboration Cycle Staging Reviews”, a “shareable HTML findings export”, and explicit guidance “for headings in screenshot-to-html skill to be not programmatic” — a small admission that some styling should not be inferred. It also records “Metrics tests: Fixes silent pass on empty chart check”, exactly the class of bug that makes agent output untrustworthy: a check that passes because it checked nothing.
The substrate held steady: tokens
The token layer did not move this week. No new specification version was published in the window. The current document at designtokens.org is a “Final Community Group Report” dated 28 October 2025, and it is explicitly “not a W3C Standard”.
Community Group activity continued without a release: a participant joined on October 28, PR #447 added positive fixtures for the color hex fallback on October 25, and PR #446 added Veritheme to the tools list the same day (W3C Design Tokens CG activity).
The portability wins landed in tools instead. Penpot 2.19, released October 28 and therefore context, lets teams “Share design tokens from a library — Store your tokens in one library and use them in every file connected to it. Tokens stay in sync with library updates”, the same guarantee components and colors already had.
The week’s agent-readability scorecard
Every in-window item, measured against the five sub-questions.
| Item (linked source) | Date | What changed | Sub-question | Verifiable signal | Gate or caveat |
|---|---|---|---|---|---|
| Figma shaders + code layers (release notes) | Nov 2 | Shaders export “up to 4096px max resolution” and into code layers; GA to all users | Resolution | Renders in a code layer set to “Play” | “Chrome version 157, rolling out this week”; code layers in open beta |
| Weave glossary (Figma blog) | Oct 31 | Defines Node, Workflow, Branch, Weave tool; side-by-side model comparison | Perception | “Each version stays connected to the steps that produced it” | Explainer, not a spec |
| Internal Weave publishing (release notes) | Oct 30 | Org-scoped Weave tools surface in the “Tools tab in Figma” | Tools | Visible “alongside Weave tools built by Figma and the Figma Community” | Organization and Enterprise only |
| Google Drive connector (release notes, help) | Oct 30 | Drive context for agent and Make; connectors on all plans | Perception | Agent names the connector and tool, then asks consent | 16 MB ceiling; per-seat authentication |
| APAC interview (Seoul Economic Daily) | Oct 31 | Quality-lags-speed diagnosis plus usage survey | Feedback | “62% said they still need six to 10 rounds of revisions” | All figures vendor-supplied, no methodology |
| A2UI + CopilotKit guide (PPC Land) | Oct 31 | Agents answer with UI via A2UI data over AG-UI | Criteria | “An agent can only assemble components the application has registered in advance” | No adoption figures or named deployment |
| Design Systems News 222 (source) | fetched Nov 2 | Three-stage token linting; round-trip spec testing; readiness checker | Criteria | 0 unresolved aliases; 0 invented properties in the diff | DS ARC is a sponsored item |
| VA.gov Design System (source) | Oct 30 | Staging-review skill, HTML findings export, metrics-test fix | Tools | “Fixes silent pass on empty chart check” | Component releases pre-date the window |
The readout is lopsided. Perception got cheap and universal: connectors on all plans including Starter, and the 4096px export path generally available. Criteria stayed do-it-yourself: the strongest computable checks this week came from a newsletter and a community write-up, not a product surface. Publishing stayed plan-gated. The loop shortened unevenly — fastest at the input end, slowest at the gate end.
FAQ
What exactly are Weave nodes, workflows, branches, and tools? Per the Figma blog: a Node is “An individual building block of a workflow on the Weave canvas”; a Workflow is “Connected nodes that turn your starting material into a finished result”; a Branch is “A separate path in a workflow for exploring a different direction from the same starting point”; and a Weave tool is “A workflow packaged so someone can run it without working with the underlying nodes.”
Can an external AI agent call Figma Weave’s image and video models? Yes. Per the October 26 release notes — context, just before the window — all Weave video and image models are accessible via the Figma MCP, “models like Seedance, Kling, and Nano Banana”, and generation “will consume Weave credits” (Figma release notes).
Is the Design Tokens format a web standard now? No. The document at designtokens.org is a “Final Community Group Report” dated 28 October 2025 and is explicitly “not a W3C Standard”. No new version was published in this window.
What stops an A2UI agent from inventing UI? Only components the application has pre-registered can be assembled — “the core of its security model”, per PPC Land. Interfaces travel as data over AG-UI. A2UI v0.9.1 is the current production release; v1.0 is a release candidate.
Which Figma plans get connectors, and which get internal Weave tool publishing? Connectors: “Connectors are now available on all Figma plans, including Starter.” Internal Weave publishing: “Internal publishing is only available on Organization and Enterprise plans.” Both statements come from Figma’s release notes.
How can I test whether my design system is agent-readable without buying anything? Run the three gates yourself: the three-stage token lint and the round-trip spec diff from Design Systems News, Issue 222, and a registry-admission check if you are on A2UI (PPC Land).
The Bottom Line
What shortened the loop this week: connectors on every plan including Starter, which makes context acquisition a per-seat action rather than a procurement project; the 4096px shader export path, generally available, which gives downstream code something faithful to consume; and the three gates, which cost nothing and convert agent error into a diff. What did not: internal Weave publishing, restricted to Organization and Enterprise, the clearest lock-in tell of the window; and the browser gate on shader-in-code-layer, which is “Chrome version 157, rolling out this week” and nothing more definite than that.
The six-to-10-rounds figure is a Figma survey as reported by the Seoul Economic Daily, vendor-supplied, with no methodology published. The argument does not depend on it being precise. It depends on the round being the unit worth measuring — and on this week’s most useful checks being the ones nobody was selling.
How This Guide Was Built
All sources in this digest were fetched with Firecrawl on 2026-11-02. Fourteen numbered sources were assembled into the working dossier; one fetch failed — the Calcite 52 release notes, blocked by antibot — and was excluded rather than reconstructed. No tools described here were run hands-on, and nothing in this article is a first-person product impression. Vendor-supplied figures are labelled at the point of use and attributed as a Figma survey, as reported by the Seoul Economic Daily, with no methodology claimed. The sponsored item from zeroheight is labelled at the point of use. The window is October 30 – November 2, 2026; items dated October 26–29 are flagged as context and never dated inside the window.