What comes after chat

Notes on the missing interaction layer between people, their context, and the models they reach for.

For about three years the dominant pattern for working with an AI was: open a chat window, paste in everything the model needed to know, and start over the next day. The chat solved a real problem — it gave us a place to talk to the thing — but it also quietly assumed something that turned out to be false. It assumed the conversation was the point.

The conversation was never the point. The conversation was the surface. Underneath, what people actually wanted was for the system to be somewhere — to know which document they were reading, which decision they were stuck on, which sentence they had just highlighted. The chat window made the model portable and the context disposable.

The friction is invisible because it's distributed

It is hard to notice this friction because it is spread thin across every interaction. You open a thread. You paste the article. You write three sentences explaining what you want. You get a response. You copy it somewhere else. You close the tab. Tomorrow the model does not know any of that happened.

The architectural mistake is treating the chat as the primitive. The chat is downstream. The primitive is closer to: I want to do something, here is what I am looking at, route it. Capture intent, attach context, dispatch the work, return the result.

What the missing layer actually does

The thing that is missing is not a better model. It is not another dashboard. It is not "agents" as a category. It is a small, persistent surface that sits above whatever you are doing — your inbox, your editor, this very essay — and lets you act on it without leaving.

Four moves, in order: capture the intent. Attach the context. Route to the right capability. Return the result, somewhere useful. Repeat at the speed of thought.

"I don't want to navigate chat threads and projects. I want to click one thing, say one thing, and have it land in the right place."

Shape, not features

A small orb on the page. A panel that opens when you tap it. The panel already knows which page you are on, has already noticed what you have selected, and is offering three or four things you are likely to want to do with that context. The orb is draggable. The panel collapses with one keystroke. There is no homepage, no onboarding, no settings carousel — the surface is the product.

The shape is small on purpose. Everything that this layer eventually becomes — voice capture, MCP routing, multi-thread dispatch, ambient memory — is downstream of getting this shape right first. The mistake to avoid is letting the layer grow before it has earned the weight.

Why this is substrate, not product

It does not care which model is on the other end. It does not care which app you are inside. It is the connective tissue between the user's cognition and the runtime — a thin coordination layer that turns ambient intent into structured execution.

Built right, it disappears. Built wrong, it becomes another chat app. The difference is whether you can feel it watching the page and whether what comes out the other end is shaped by the place you were when you spoke.

What's next

The standalone prototype proves the interaction model on a single page. The browser extension lifts the same surface onto any page on the web. The developer SDK lets a host application define its own context hooks and action routes.

If it works, you will stop noticing it, which is the highest compliment a substrate can earn.