Exploration of home screen with modular tiles

This design explores a home screen that is comprised of a modular set of cards.

While the input UI hints at the 'tool belt' it is fundamentally familiar from other agent harnesses. However the main part of the screen is not occupied by chat, but rather a modular set of cards. These could be the result of the Satellite agent transforming user input into an artifact.

• Imagine a voice memo that gets transcribed and automatically shared based on the topic it was associated with.
• An item that is phrased as a todo and gets auto-categorised alongside a subtle transformation of the UI into a todo item with a checkbox.

Beyond those simple components that combine agentic output with generative UI, the user and the system can collaborate to build fully tailored interfaces based on the information the user adds to Satellite.

If the user is collecting types of beers, taking a photo of a newly tried one can transform the information from the photo into a simplified card that takes on subtle cues from the brand of the bottle. The bottle itself is normalised (by removing the background and normalising the bottle using an image generation model), the typography is automatically selected by the model from Google's font library to be the most suitable option for the topic at hand.

This initially generated user interface can be adapted by the user for this specific category and any adaptation will apply to any other item of the same category. If the user prefers to remove the nutritional information but keep the country of origin, the user can simply instruct Satellite to do so.

As simple categories such as 'Beer', 'Todos' and 'Movies' do not fully express the depth of the context graph, the category pills associated with each card can expose a second level of depth. You can see this in 'Movies', where the system chose to expose that it understood based on the user's context that 'If I had legs I'd kick you' is part of a collection of movies the user had already seen, rather than a watch-list the user might keep as well.

While this exploration of Satellite is intended to be a mostly self-organising 'stream of consciousness'-style timeline of any piece of information the user (or the model via MCP) decides to capture, the top shows a hint of how filtering information might work.

Swipe gestures (or taps) let the user filter the Satellite stream by the different categories. The leftmost category is the user themselves. Swiping to the left reveals any truly personal information Satellite remembers. These could be certain preferences, ID numbers, potentially medical records. It is simultaneously a new, AI-first take on 'Settings': Rather than heavy, rigid UI — the users preferences are essentially just a special collection of memory.