Skip to content

Blog

Google A2UI Express, Figma Weave and its editing model, Webflow DevLink and Amazon Alexa Home Map. Five official announcements from this week, checked against the originals.

Published

On October 27 Google introduced Express, a new inference format for A2UI, on its developers blog. A2UI is a message format an AI agent uses to describe a screen, and Express reduces how much the model has to write. The model produces a compact intermediate form, and the SDK turns it into standard UI messages.

Google gives two numbers. Tokens are up to 3x fewer than with Direct JSON, and latency is up to 3x lower. When a decision model chooses the layout, the post also cites up to 80x. The decision-model format is not in the SDKs yet, and the post does not state the baseline for the latency figure. I read "up to" literally, as a ceiling. Express is available in the Python, TypeScript and Dart SDKs.

AI modelcompact Express form
SDKstandard UI messages
Rendererthe screen code you already ship
  • AI model → SDK
  • SDK → Renderer

Image: Google Developers Blog

The Figma post dated October 31 is a how-to guide, not a product launch. You can send one prompt through several models and compare the results, branch from an image you like to try other prompts and edits, and save the order of your steps so a teammate can run it again with new inputs. A saved workflow can be packaged as a Weave tool that others run without touching the nodes.

It suits teams that repeat the same job every week, such as product photos or campaign images. The point is less one good result than keeping the path that led to it.

Image: Figma Blog

The Figma release notes carry one more Weave item. The entry dated October 30 describes publishing Weave tools only inside an organization. Tools published internally appear in the Tools tab, on Organization and Enterprise plans. My reading is that a team can turn a repeated job into a tool and run it as a shared asset, but because of the plan limits it is not open to every user. The source is the Figma release notes.

Figma described how it trained its own model for editing designs that already exist on the canvas. The jobs are everyday ones, such as adjusting spacing, changing colors or reworking a layout after feedback. Designs were fed to the model in Figma's native file format, and the training requests included theme and component changes to design system assets. Work such as new screen generation and vectors goes to other companies' models.

  • Task speed
    20%

    Typical tasks got faster from beta to general availability

  • Undo rate-60%
    60%

    Users undid the agent's work less often

  • Against the old agent
    62%

    Share of expert comparisons where the new agent matched or beat the old one

  • User ratings+33%
    33%

    User ratings were higher

Source: Figma Blog. These are Figma's own evaluation figures, not independently verified.

The post does not spell out the baseline or the period. So reading what Figma chooses to measure is more useful than reading how much better it got. Tracking task time and undo counts together suggests that fixing things without mistakes counts as performance, not only speed.

The Webflow CLI command webflow devlink export exports Webflow components to a React project. In the other direction, webflow devlink import bundles a React library and uploads it to your workspace. Export needs @webflow/webflow-cli v1.19.0 or later and Node.js v22 or higher, and a project that uses the exported components needs React v18 or higher.

The docs also cover keeping styles from leaking. With cssScopes on, component styles are wrapped in CSS @scope, and it is on by default. Only components can be exported, so ordinary elements on the canvas have to be turned into components first.

Webflow components
React projectwebflow devlink export
  • Webflow components → React project
React library
Webflow workspacewebflow devlink import
  • React library → Webflow workspace

Image: Webflow developer docs

Alexa Home Map, as Amazon's newsroom describes it, shows your home and property as an interactive 3D map and places each device where it actually sits. You can zoom into rooms and switch floors, and the camera timeline rewinds the last 24 hours. Light color and brightness, the music that is playing, day and night, and the weather are reflected on the map.

The preview starts later this year on the iOS app, the Alexa and Ring apps, larger Echo Show screens and new Alexa Tablets. Some features need Alexa+ or a paid Ring subscription. There is no exact date yet.

Image: Amazon newsroom

If I had to tie the five items together, it would be "the design system is becoming the rulebook for AI." Google has the AI write lighter UI messages for the renderers you already use, and Figma trains a model fitted to design edits. Both narrow the format and the scope of the work before giving the AI more freedom. That is my interpretation, not a fact.

Webflow opened a two-way road between design and code, and Alexa shows status as place and movement instead of a list. If I had to try one, I would start with A2UI Express. The structure, where the AI writes less and the SDK handles the rest, will probably outlast the speed numbers. Publishing a component registry and llms.txt, as dsui does, points the same way: put the list of parts an AI can use out in the open first.

One item was left out. The report that KT M Mobile's eSIM service mobi won a Red Dot Design Award is dated July 24, so it does not belong in this week's news.