Published November 18, 2026

A2UI 1.0 Ships Alongside a2ui_flutter: Generative UI Gets Guardrails, Speed, and a New Name

Generative UI has spent a year in the awkward phase: impressive in demos, fragile in production. The Flutter team's answer arrived on November 17 with A2UI 1.0 and a2ui_flutter, the package that until now shipped under the name genui. This is the first version that reads like something you could ship to paying users without a list of caveats.

The premise has not changed. An agent that reasons over a user's data is a poor fit for a chat log full of numbers and follow up questions. Generative UI lets the model answer with Flutter widgets instead. What changed is that the protocol and the Dart client now share a name, a version number, and a set of structural limits on what a model is allowed to build.

Where Generated UI Still Loses

The launch post opens with a rule of thumb that is worth repeating because it is the question most teams get wrong. Generative UI is not a replacement for deterministic screens. Forms, lists, and navigation tabs are fast, predictable, and already built. If a static form handles a workflow cleanly, generating it is a downgrade with extra latency and a new failure mode.

The case for generation is narrower: workflows where the right interface depends on reasoning over open ended context. The team's Commis demo, a catering app for jobs and recipes, is the clearest example. Ask whether a menu is safe for a host with a shellfish allergy and the agent confirms the answer, then renders a modified menu containing only safe recipes. There is no allergy checkbox and no ingredient filter in the product.

A Smaller Architecture With Five New Capabilities

A2UI 1.0 collapses six concepts from the genui era into three. The transport abstraction is gone, so the package no longer dictates how messages reach your app. MessageProcessor replaces the SurfaceController and Conversation pair, taking over catalog registration, message validation, and surface state in one place.

The first addition is a pair of schema properties called allowedParents and allowedChildren. Give a model too much freedom and it will notice that a Button accepts a child and decide to nest a DataTable inside one. The allowlists appear in the catalog schema inside the model's prompt, so the model knows the nesting rules before it generates anything, and the client enforces them again before building widgets. A model that ignores the prompt gets an UNALLOWED_PARENT or UNALLOWED_CHILD error instead of a crash for the user.

The second is the Express inference format. Every token emitted during a streaming render is latency the user feels, and JSON spends much of that budget on quotes, braces, and repeated keys. Express expresses the same tree in a compact function call syntax, describing five components in 46 percent fewer characters than minified JSON.

Solid Software measured the effect in GenUI Clinic, a Flutter healthcare prototype: full screen dashboard generation dropped from ten seconds to five after switching to Express.

The third is multiple catalogs per surface. Under A2UI 0.9 a surface was bound to a single catalog, forcing teams to merge core components with feature widgets by hand. Now MessageProcessor takes a list containing a CommonCatalog for shared layout primitives, a domain catalog owned by a feature team, and a separate chart catalog. The client resolves component names across every registered catalog.

The fourth addresses a frustration in every prototype. Tapping a chart bar or a suggestion chip deserves a new model turn. Dragging a slider or stepping a counter does not, yet the old design charged every interaction the same round trip. callAgentFunction lets a rendered widget invoke a deterministic function the agent backend exposes, so the backend recomputes the result without involving the model and returns data into the existing component.

Somnio Software built their open source portfolio companion around exactly that split. They filtered eight candidate interfaces down to three where generation beat a static form, then kept every calculation deterministic. Drafting a ten thousand dollar order across BTC, ETH, and SOL and tapping to adjust leg quantities updates weights, fees, and remaining cash in place in under a second.

The fifth is extensibility metadata, which carries structured out of band data on surfaces, components, catalog definitions, and actions. Routing flags, styling hints, and correlation tokens that tie backend logs to mobile telemetry travel with the message instead of being smuggled through widget properties.

Agents in Three Languages

a2ui_flutter only renders. The other half of the stack is an agent SDK that teaches a model your catalog, compiles its output including Express into A2UI messages, and validates the result before it reaches the app. Dart is now a first class option through a2ui_agent, joining a2ui-agent-sdk for Python, a2ui-adk for Google's Agent Development Kit, a2ui-a2a for the A2A protocol, and @a2ui/agent for Node.js.

The Dart path is the one to watch. Pair a2ui_agent with Genkit Dart and the agent, its tools, and the Flutter interface are written in a single language. The genkit_a2ui middleware, still in preview, builds on a2ui_agent to wire that together. Very Good Ventures used the combination for an open source smart home panel that composes a different dashboard for each family member, with a Dart Frog backend running the agent and Gemini Live handling speech.

Teams with a genui integration can move straight to a2ui_flutter, since the concepts map cleanly even though the controller pair is gone.

The payoff is a stack where the failure modes are defined. A model that proposes an impossible hierarchy gets a validation error instead of a crash. A model that generates too many tokens gets a format that produces fewer of them. A widget that needs to do arithmetic asks the backend instead of the model. Those are the properties that separate a demo from a product.