The missing interface

A wave of personal AI agents shipped this year, and nearly all of them landed in the same place: a message thread. OpenClaw turned the chat apps people already use into a command center for their computer, Instinct launched as an assistant you text or call, Meta shipped Muse with a name and a cute avatar, and a few weeks later OpenAI brought dots into ChatGPT.

It has been a while coming. Most of the interesting interface work of the last two years went to developers, while everyone else got a text box. The gold rush was agentic coding, and the consumer side sat still.

Underneath, most of these are the same idea: give the AI its own computer, let it use the browser for you, and talk to it through a thread. OpenClaw proved it but stayed too technical for most people. Muse and dots wrapped it in a better harness and an interface people can relate to. Instinct went a slightly different way, an assistant with its own phone number and email address that you message on WhatsApp. The real difference is where they live. Muse is its own app, dots sit inside a bigger one, and Instinct has no app at all.

Meta’s version is the one I find most interesting, mostly for what it reused. Twenty years of learning how people behave around a name and a cute avatar at the top of a thread. It went straight to the top of the App Store.

What I keep wondering is why the result is so plain. If an agent browses the web for me, why does it come back with a bulleted list? Take Airbnb. I like its interface; the problem is the time it takes to pick, and how few options match what I want. An agent that searched Airbnb and its competitors and came back with five options, in something that looks and feels like Airbnb, is something I would use every week. Some days I would just tell it to surprise me and book. But that missing interface in the middle feels like the biggest gap in consumer AI right now.

Generative UI is the part I think gets too little attention. Most of the talk stays at the extremes: everything is a chat, or everything is an app. The interesting work sits in between, building the interface you need from trusted building blocks. Funny enough, the lab I feel is making the most progress here is the most enterprise-focused one: Anthropic’s artifacts already turn an answer into a small app you can share. The usual worry is an interface that changes every time and that you never learn. I think it could be the opposite: an interface that learns you. Feeds did a version of this, Netflix too, but this is the whole interface shaped around you as you go, something game makers have dreamt about for years.

That is why I am watching Apple most closely. They are the one player that doesn’t need a thread at all. They own the screen, a design language people already know, the context in your messages, mail, and calendar, and a way for apps to tell the system what they can do. The new Siri can already read that context and what is on screen, but for now it mostly answers questions. I think they have everything they need to make something different.

I spent the better part of a decade making interfaces before I made products. This is the first stretch in a long time where that question feels properly open again.