Matt Stokes · Research Synthesis
The Future of the Interface
Project Solara and generative UI research point toward interfaces that move with the task while task state and authority remain inspectable across devices.
Microsoft describes Project Solara as a platform for agent-first devices. Its interface concept, just-in-time UI, sits between responsive layouts and fully generated experiences. The same agent can adapt across a badge, a desk device, or a larger display while keeping enough structure to remain consistent.
Google Research is testing a more generative end of that range. Its generative UI work creates a custom interactive experience for a prompt. In its evaluation, people preferred those interfaces to standard model output when generation time was excluded. Human-designed sites still ranked highest.
Together, those results support a product that can change its layout for a badge or a desk while preserving the rules that govern what the agent may do.
The interface moves from apps to tasks
An app gives the interface a fixed boundary. It owns its data, controls its actions, and presents a known set of screens.
An agent can cross those boundaries. It may gather information in one service, prepare work in another, and return later on a different device. A generated surface can show the right controls for the moment. Someone who picks the work up later still needs the goal, the current status, the source material, the pending decisions, and the actions already taken. That record has to live outside the current layout. If it is baked into a particular screen, a badge, a desk device, or tomorrow's session can drop the work they came back to finish.
Authority becomes the stable interface object
When an interface can redraw itself, position and visual hierarchy stop being reliable ways to communicate control. What the agent may sense, which memory it can use, which tools it can call, and where it must stop for approval have to live underneath the rendering. Otherwise a new layout can hide a grant of power that was obvious on the previous screen.
A badge has almost no room, and it still has to disclose a large grant of authority. On a voice surface there is no screen left to hold the warning, so the response itself has to name a pending commitment. If adaptation changes those answers, the person can no longer tell what the system is allowed to do.
Generated UI needs a controlled vocabulary
Fully open generation offers range. It also makes failure hard to test, because a broken state can mean anything. Solara's middle position is useful because the agent adapts inside known content types and interaction patterns. The generated surface should be composed from a small set of those patterns: a status the person can read, the evidence in play, a choice that still waits, an approval that can stop the work, a progress state, and a receipt for what already happened. The agent can arrange those pieces around the task. Each one keeps a known meaning, an accessibility contract, and a failure state.
Layout and media can still be generated around that set. Designers then have something concrete to test: whether people understand the current state, and whether they can stop or correct the system.
The commit boundary should not move
A generated interface may simplify routine work. As an action becomes harder to reverse, the same interface has to become more explicit.
Sending a draft, changing a reservation, or publishing a decision should appear as a clear event. The person needs to see what will happen, why the system is proposing it, and whether the action can be undone.
When the next action is uncertain or hard to undo, the surface has to name the pending commit and wait. Independence can increase only inside a grant the person can open and revoke.
Agent-first devices widen the audience
Solara also moves agents into shared physical space. A badge or room device can sense people who never chose to use it, so recording state needs a physical signal, and sensor controls have to stay visible to more than the operator. The operator's permission cannot silently stand in for everyone nearby.
Sources
- Microsoft. Composing a new platform for agent-first devices, 2026.
- Google Research. Generative UI: A rich, custom, visual interactive user experience for any prompt, 2025.
- Google Research. Self-Evolving Systems: Moving Beyond Deterministic Interfaces to Adaptive Generative Interfaces, 2026.
The longer Long-Term User Experience of Agentic Systems working paper develops the authority model and interaction grammar in full.