My MacBook Air gave up at about 4 or 5 development agents working at the same time. The CPU was the first sign.
A laptop's job is to move around with me. It closes in my bag and changes networks, and each time it does, the agents working on it stop.
So I moved the work to a dedicated machine that stays on, and the Mac became the front end I use to direct and review it. Agents made the desktop relevant again.
That machine is where the work lives now. The agents keep running when I close the laptop, and their memory stays with them.
Those agents work in terminals. The next ones will use the whole computer.
Give the agent its own computer
Both of this month's new frontier models, OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1, can use a computer the way a person does: read the screen, click, type and open applications.
I want the agent to use the computer, open its own software and work in it around the clock, with its own identity when it needs one. That's my golden standard: a full computer for the agent.
The more an agent is allowed to do on a machine, the more that machine should be its own.
Anthropic recommends "a dedicated virtual machine or container with minimal privileges" for computer use. OpenAI advises: "Use an isolated browser or VM". Microsoft gives each agent its own identity and says agents "should operate in their own secure, computing environments".
For me, computer use is what a Mac mini is for. With its own computer and its own logins, that's the agent I'd hand the sensitive work.
Another reason: we already hand AI a lot
People ask AI the questions they'd take to a lawyer or a doctor, and upload bank statements and medical records to get a better answer. At work it's the code and the terms of the deal.
More and more of what we give it is confidential. Tell a lawyer, and privilege can protect it. This year a US court ruled that a defendant's conversations with Claude weren't protected by attorney-client privilege.
This month OpenAI was asked whether users' sessions had helped one of its models. It said: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." The consumer terms of OpenAI, Anthropic and Google all allow de-identified data to be used to improve their services.
De-identified data keeps the idea. For secret research or a confidential deal, the idea is what's worth protecting. With the new hardware and local models, keeping that step on my own machine becomes a routing decision I can make task by task.
And availability, the reason I started with, applies to everyone. Many AI users say they now do work they couldn't a year ago, and on 3 September ChatGPT, Claude and Grok went down at the same time for a few hours. If your day runs through them, it stops with them.
Confidentiality and availability both give me a reason to run AI on a machine of my own.
What to do about it
My answer: keep the agents, their memory and the confidential work on a machine I control, and decide what leaves it.
Open-weight models make that possible. A capable model now fits on a desktop GPU or a Mac, under a licence a company can accept.
The companies that make the hardware and the agents are building for exactly this. Perplexity started with an agent in the cloud, moved it onto an always-on Mac mini, and in August released a version that runs the models on your own NVIDIA GPU and asks before a step goes to the cloud.
Apple is positioning the new Mac mini for "always-on, deskside agentic computing" and the Mac Studio to "run enormous LLMs entirely on device with complete privacy". Perplexity launched Portable Computer on NVIDIA's desktop DGX Spark, and AMD is pitching its Ryzen AI Max chips for running models locally.
What I get is the choice. Each step can run on my own machine or go to a hosted model. The local model can propose which, and I decide.
What leaves your machine today, and who decided it should?