The state of generative UI, November 2026
Generative UI went from demo to default this year. The protocols consolidated into plumbing, the SDKs churned, and the one layer nobody standardised is the one that decides whether people get hurt. A field report with dates.
Twelve months ago, "generative UI" mostly meant a weather card in a demo. This month, ChatGPT answers ordinary questions with charts and calculators on every tier, Google renders generated interfaces in Search, and the same embedded app can run inside Claude, ChatGPT, VS Code, Microsoft 365 Copilot and Gemini. Here's what changed, what consolidated, what broke, and what's still missing — with dates, so you can check us.
1. It became the default answer format
The pattern of the year is that the largest AI products stopped answering in paragraphs when a screen would do better.
- Google shipped generative UI with Gemini 3 in November 2025 — dynamic views and visual layouts in the Gemini app and Search's AI Mode. Its research paper on the system reports generated pages preferred over markdown answers 82.8% of the time.
- MCP Apps became the official standard for interactive UI inside AI hosts on 2026-01-26, proposed two months earlier by MCP maintainers at Anthropic and OpenAI together with the creators of the community MCP-UI project.
- ChatGPT made generated, interactive answers ("Intelligent UI") standard in October 2026.
- Gemini and Search started rendering MCP Apps at the end of October.
The practical consequence: if your product ships an MCP App, it can now appear inside most of the AI products your customers use. And it will be judged next to the hosts' own generated interfaces.
2. The protocols consolidated into plumbing
A year ago the question was "which protocol wins?". The answer turned out to be: they sit on different layers, and they now wrap each other.
| Layer | Standard | Status, Nov 2026 |
|---|---|---|
| Agent ↔ frontend events | AG-UI | 1.0 in September 2026: a normative spec, a JSON Schema, a conformance corpus, interrupts and resume, subagents, reasoning events |
| UI described as data | A2UI | v1.0 current since 2026-11-06; renderers for web, Flutter, SwiftUI and Jetpack Compose |
| UI described as data | json-render | 0.22; ~19k GitHub stars in ten months; renders to web, mobile, PDF, email, video, terminal and chat cards |
| UI described as data | OpenUI Lang | 1.0 in October 2026 (Thesys, open source) |
| UI hosted in AI products | MCP Apps | stable since January; rendered by every major host |
| Pages exposing tools to agents | WebMCP | W3C community-group draft; origin trials in Chrome 149 and Edge 150 |
AG-UI now carries A2UI surfaces and MCP Apps as activity events; A2UI has a component for embedding MCP Apps; json-render ships an MCP Apps server. A mature product uses three of these at once. The fight people predicted didn't happen.
What did happen is version skew. A2UI changed its message envelope several times in a year — nested to flat components, path to @path bindings, "client/server" to "renderer/agent", a new MIME type. Integrations lag: in November, first-party middleware still prompts older shapes. MCP core went stateless in July and deprecated sampling. WebMCP moved from navigator to document in May. If you integrate any of these, check the in-band version and reject what you don't speak.
3. The money went elsewhere
The protocols are free and the downloads are large — the MCP Apps SDK alone sees millions of weekly installs — but the money pooled in platforms and app generators, not generative-UI SDKs. CopilotKit raised a $27M Series A in May. Tambo shut down its hosted cloud this autumn. Thesys turned its C1 API into the open-source OpenUI project. OpenAI is winding down Agent Builder at the end of November. LangChain archived its reference chat UIs. Vercel's AI Elements is folding into shadcn/ui.
Lesson for builders: pick standards over vendors, and keep your conversation data in your own database.
4. The SDKs churned hard
If you learned generative UI from a tutorial in 2025, much of it is now deprecated:
- The Vercel AI SDK went from v5 to v7 (June 2026).
addToolResultis nowaddToolOutput; approvals moved totoolApproval, and can be signed so they bind to an exact call;MCPAppRendererlets you render MCP Apps in your own product. - CopilotKit deprecated every v1 hook in August; the v2 API is
useFrontendTool,useHumanInTheLoop,useRenderTool,useAgent,useAgentContextanduseInterrupt. - assistant-ui replaced its tool-UI API with
defineToolkit. - LangGraph shipped a new React client and renamed its platform.
The ecosystem's half-life is roughly six months. The defence is architectural: put your own tool and component registry between your product and any framework, so the definitions survive a swap. And watch your coding agent — they generate deprecated APIs by default.
5. The one layer nobody standardised
Read every spec side by side and one gap stands out. No protocol defines what it means for a generated interface to have the person's consent before it does something irreversible.
- MCP Apps makes sandboxing mandatory, but asking the person before a UI-initiated tool call is only a "MAY".
- MCP's own maintainers wrote in March that tool risk annotations are untrusted and that no major client shows them in approval prompts.
- WebMCP's draft admits that sites can misrepresent their tools' read-only and consequential hints, and that there's no way to verify them.
- The effect-safety vocabulary is fragmented: MCP has destructive hints, WebMCP has its own hints, A2UI has user-activation requirements, json-render has a
confirmbinding, AG-UI has interrupts, the AI SDK has approval requests. None defines reversibility.
Meanwhile the incidents kept coming: an MCP-UI vulnerability that let malicious UI call tools without confirmation, agents deleting production databases, agents paying on convincing fake shops. And the research says models are no help by default — a CHI 2026 study found a majority of model-generated web components contained at least one deceptive design pattern.
Evaluation hasn't caught up either. Benchmarks climbed from pixels to interactions to functional tests run by agents, but the next rung — consequences — is missing. Nothing widely used asks whether a generated interface showed the right confirmation before an irreversible effect, whether that effect can be undone, or whether the button's label matches what it does.
6. What we expect next
- Interfaces become disposable; their contracts don't. Catalogs, AI-facing design-system specs and tool contracts will be versioned like APIs.
- App-store review gets real. There are now four review gates for embedded AI apps — OpenAI, Claude, Microsoft 365 and Google — and they behave differently. Cross-host conformance will become a job.
- Every interface gets a second audience. Agents will operate the same surfaces people do, and the components that win will expose one implementation to both, with declared consequences.
- Trust moves from review to evidence. Nobody can review every generated screen. The question becomes whether you can prove what a screen does, and that a person agreed before anything irreversible happened.
That last one is the problem our lab works on. We think the field needs an open, testable Generated UI Safety Profile — consent before destructive effects, labels that match effects, approvals bound to arguments, declared egress, operable consent dialogs — mapped onto MCP Apps, AG-UI and A2UI. We'll publish a draft, and a benchmark that grades what generated UI does, not how it looks. If you build or host generative UI and want to shape it, get in touch.
Sources for every claim above are in the Field Guide's research notes. For the full protocol picture, start with Module 02 · The protocol stack; for the reasoning behind the safety argument, read The model only proposes.