The next operating system

I think most homes will have an AI workstation. A box in a cupboard, without a screen, running a new kind of operating system. Phones, laptops, and whatever we end up wearing on our faces will be thin clients that connect to it.

I already live a version of this. A Mac Studio in my home runs nothing but an AI agent. Nobody looks at its screen. I reach it from my laptop and get the results back as text, files, or pages.

Today’s operating systems exist to run apps that someone else built and you installed. I don’t think that survives.

I’ve argued before that the value of software is racing to zero. You no longer ask for a tool. You ask for a job to be done, and the machine writes whatever software the job needs along the way.

If that’s true, you stop installing apps. The software for a job gets built when the job starts and thrown away when it’s done.

What you keep are workflows: the routines and processes for jobs you do again and again, written down so the machine can repeat them. The monthly close, the weekly report, the way you onboard a new client. Workflows are the new apps. So are MCP servers, the connectors that let an agent read your calendar or pay an invoice. You set them up once, and every job can use them.

So the new operating system doesn’t manage windows and files. It manages your context: your documents, your history, your preferences, and the credentials to act on your behalf. The interface is a conversation, and the screen is whichever one you happen to be holding.

Why a separate box? If the AI runs your books, it can’t stop when you close your laptop. Invoices arrive at night, and payments fall due at weekends. The agent needs a machine that is always on.

That machine could live in the cloud. I’d rather it lived at home. Your context is the most private data you have, and I want it on hardware I own. An agent that works all day is also cheaper on a box you’ve paid for than on compute you rent.

There is a tension here. In the same note, I argued that compute is one of the few things that keeps its value. If compute is scarce, the biggest models will stay in the data centres that can afford them. I don’t think that breaks the idea. The box at home doesn’t need to run the smartest model. It holds your context, and its local model orchestrates the work and handles anything sensitive. Complex work goes to frontier models in the cloud.

It looks a lot like the mainframe era, when terminals connected to one central computer. This time, the central computer is yours.