Why does an AI agent need its own computer?
A chat window forgets and a laptop sleeps. An agent that keeps logins, files, a browser and a schedule needs a machine that is always there and nobody else uses: the four things a computer gives it, and why a shared one is the wrong shape.
Because the useful part of an agent is not the model, which you can rent by the token from anywhere, but everything the model needs around it: a browser that stays signed in, tools it installed last week, files it is working on, a memory that accumulates, and a clock that keeps going when yours stops. A chat window has none of that. Your laptop has all of it but sleeps, travels and holds your own life. A computer of its own gives the agent four things at once: always-on time, persistent state, a place to act, and a boundary around what it can touch.
The model is the cheap part
Every AI agent is a loop: read the situation, decide, use a tool, look at the result, repeat. The deciding happens in a model that lives in a provider’s datacenter and costs fractions of a cent per call. The tools are where the value is, and tools need a place to run. A browser needs an operating system, a display and a profile to keep its cookies in. A script needs an interpreter and its packages. A file needs a disk that is still there tomorrow. Take the computer away and you have a very articulate program that cannot do anything, which is what a chat window is.
Four things a computer gives an agent
Time. A morning briefing at 06:00 needs a machine awake at 06:00. A price watch that checks every hour needs a machine that exists every hour. An agent that answers a message from your phone at midnight needs to be somewhere when the message lands. Laptops close; phones lock; the agent’s computer does neither.
State. This is the one people underestimate. An agent that is signed into your calendar, has a package installed for reading PDFs, keeps a folder of the research it did last month, and has learned that you prefer short answers is an agent that gets more useful every week. Each of those is a file or a session on a disk. When the disk is not the agent’s own, the state has to be rebuilt. One developer’s first day with a coding agent that got a fresh machine per task: “the main issue then becomes that you have a fresh machine per task; which doesn’t have any of your env or access to tools or clis you use.” He spent hours on setup scripts to recreate, per task, what a persistent machine simply keeps.
A place to act. Filling in a form, buying a ticket, moving a file, running a report: these are actions in a browser or a shell, and both are programs that need an operating system under them. The tool an agent uses most is a computer, in the ordinary sense.
A boundary. What the agent can reach ends where its machine ends. On its own computer, that is its own files, its own logins, its own network. On yours, it is everything you have.
Why not the laptop you already have
You can, and it is a fine way to try one. Three things push people off it within weeks. The laptop sleeps, so the agent is awake only when you are. It roams: the home network, a cafe, a hotel, each with its own address and rules. And it is yours, in the sense that matters for security: the machine an agent runs on is the boundary of what a mistake can reach, and a laptop with your password manager, your browser sessions and your tax returns is a large boundary to hand to software that reads untrusted web pages for a living. The practical version of the same point came from developers running several agents on one machine: worktrees and port registries to keep them apart are, as one put it, “lipstick on a pig. The solution is VMs.”
Why not a shared computer
Some products give you several agents on one machine. That can be efficient, and it has a cost the vendors state themselves. Grok Bot’s docs, on the single computer every bot on an account shares: “Do not use separate Bots as a security boundary.” Files, browser cookies, credentials and the command line are visible to every agent on it. The strict shape, one agent per machine, is also the simplest to reason about: whatever one agent can do, only that agent can do.
What “its own computer” has to include
Here is the test that separates a computer from something marketed as one: what survives an update? On a laptop, a restart loses nothing. Grok Bot’s docs are candid that on its computer, files in /workspace, browser state and sign-ins survive, while “manually installed packages” are replaceable. A user’s field report translated that into practice: “Plan to reinstall Tailscale, SSH tooling, and anything installed with apt.” That is a workspace, not a computer, and it matters because the tools an agent installs for itself are part of its accumulated competence. When you choose where an agent lives, ask whether the whole home persists or one folder does.
The Mac mini instinct, and its cloud equivalent
The reason so many people buy a Mac mini for their agent is that it answers all four needs at once: it is always on, keeps everything, runs a real browser, and is not the machine they work on. Its costs are the up-front price, being tied to one address and one home network, and being yours to administer. The cloud version of the same idea is a small server that is the agent’s alone: a VPS you run yourself, or the same thing with the operations done for you. Mac mini, VPS or managed machine is the comparison; whichever you pick, the thing you are buying is the same: a computer that belongs to the agent, so that everything the agent learns and sets up belongs to it too.