Generative UI belongs on the Home Screen

The model is learning to choose the interface, not only the words. Glance lets that choice land on the iPhone widget you already see.

You wake the phone. A widget is already there. Yesterday it was one number. This morning it is three lines and a time, because that is what the morning is. You did not open a chat. You did not pick a template. An agent that knows the day chose what the slot should show, and how it should show it.

That is the shift hiding inside a phrase the labs have started to use: generative UI. The model does not only write the answer. It decides the shape of the answer. A sentence when a sentence is enough. A chart when the comparison is the point. A control when you need to change one thing and see what happens.

The interface is no longer designed once

For decades a screen was drawn before you arrived. A designer placed the button. An engineer kept it there. You learned the app, and the app stayed put.

In November 2025, Google Research described a different arrangement. A model would generate the experience itself: pages, tools, and simulations made for the prompt in front of it, rather than chosen from a catalog of apps. The same week, Gemini 3 started testing it. One experiment, visual layout, turns a request into a magazine-like page you can revise. Ask for three days in Rome and the itinerary arrives as something you can browse, then reshape. Another, dynamic view, codes a small interface for the question. The Verge tried the Rome trip and got a page that asked how fast you wanted to move and what you wanted to eat, then redrew itself. Wired watched a physics question become a custom visualization instead of a paragraph.

In March 2026 Claude began dropping charts and diagrams into the thread when the conversation called for them. In October 2026 OpenAI shipped Intelligent UI with GPT-6. The Verge’s Emma Roth described replies that arrive with pictures, charts, forms, and buttons: a bicycle you can tap through, a bill splitter, a small tool built for the question and left inside the answer. OpenAI’s own account is the useful one. They gave the model a library of native components, and they trained it to decide when to use them. A comparison sits side by side. An explanation becomes a diagram. A plain question stays plain text.

The pattern is the same in all three. The model is no longer filling a box someone else designed. It is choosing the box.

The box still lives in the chat

Open the app. Ask. Wait. Look at what the model drew. Close the app, and the drawing stays in there.

That is a real change in how a conversation looks. It is still a conversation. You have to go get it. The Home Screen does not work that way. It is the surface that is already there when the glass lights up. Widgets were built for that glance, and then they were frozen. A weather widget is a weather widget. A calendar widget is a calendar widget. The system decides the chrome, and your data fills the holes.

Generative UI, if it is going to matter outside a chat, has to survive that constraint. A few seconds of attention. A fixed rectangle. No room for a webpage, and no appetite for a new app every time the question changes.

A slot, and a tree

Glance treats the widget as a slot you pin once. Behind the slot is a layout and a feed. The layout is a JSON tree. The feed is the latest snapshot. iOS draws the tree with native views: the type, the color, the corners, the sizes Apple already gave widgets. Small, medium, large, or a full page. The agent does not ship an app. It does not ship HTML. It speaks that tree.

There are two ways the face gets onto the glass.

A saved layout is a face you want again. Revenue should look like revenue every hour. The tree stays. The numbers move. You can reuse it, and other updates can fill the same holes.

UI on Demand is the other way. The tree travels with the update. ChatGPT, Claude, or Cursor calls Glance and sends the content and the layout together. Glance checks the tree the same way it checks a layout you saved by hand. If the tree is legal, the Home Screen draws it. The next push can be a different tree. Nothing is filed away as a layout you have to manage. The slot did not spend one.

Same endpoint. Same rate limits. A different question about what is allowed to change.

One slot, three mornings

7:10. The agent sends a medium widget. “Standup in 12 min.” Three names. That is the whole face.

12:40. Same slot. A different tree. The title of a pull request, and whether it is waiting on you.

6:05. One number. The day is over, and a paragraph would be noise.

The model had the freedom to redesign the slot each time. It did not have the freedom to invent a new kind of screen. The canvas is small. It has to be understood in about two seconds. Text, stacks, a chart, a second screen you tap open: the same pieces a saved layout can use. OpenAI’s component library and Glance’s layout language are the same kind of bet. Constrain what can be drawn, and let the model decide what this moment needs.

That is the part chat cannot do. Intelligent UI is remarkable, and it remains inside ChatGPT. Gemini’s dynamic view is remarkable, and it remains inside Gemini. Glance is the first app that lets the agent you already use put a fully generated interface on the Home Screen, drawn by the iPhone, sitting there before you ask.

When the face should stay

UI on Demand is the wrong tool when the chrome is the brand. A public widget people subscribe to, a clock, a number you check every day: save the layout, and push the fields. The face stays recognizable. Only the snapshot changes.

It is also the wrong tool for a schedule Glance runs on its own. An automation keeps a saved layout, so a refresh does not redesign the widget out from under you. UI on Demand is for a push where the design is part of the message. It is on Pro and Power. The tree has to come back in full on every update, because there is no memory of a layout to fall back on. What is on the glass is exactly what the last push said the glass should be.

The catalog you already arranged

Google’s researchers wrote that people should receive an interface for the need in front of them, instead of picking an app from a catalog. The catalog is not going away. You already arranged one. It is the grid of widgets on the phone.

The next step for generative UI is to leave the conversation and take one of those squares. The model chooses what you see, and how you see it. Glance is the slot where that choice becomes native, and stays on the glass after the chat is closed.