The AI Agent Operating System I Run My Business On
My business runs on an AI agent operating system. The phrase sounds like a product you can buy. It isn’t one. It’s a Mac that stays on, Claude Code running in terminal sessions that never close, a folder of plain Markdown files holding everything the agents need to know, and an email address the agents send from and receive on.
Here’s what’s actually in it, what a normal day through it looks like, and the parts that still need me.
What the operating system is made of
A machine that stays on. A Mac Mini at home, headless, no monitor and no keyboard attached. Agent sessions run inside tmux, so they stay alive whether anything is connected or not. Close the laptop and the work carries on.
The agents. Claude Code does most of it. Hermes, Grok and Codex sit alongside for jobs where one of them fits better, and the heavier reasoning work goes to a Claude Opus seat while quicker jobs run on Sonnet. Which agent runs a job matters less than what the agent already knows when it starts, which is the next part.
A memory made of files. There’s no database in it and no retrieval pipeline. A CLAUDE.md at the top teaches the agents how the business makes money, who the clients are, how I write, and what has already been decided. Under that, a folder per client and per project holding its own state, history and open questions. Every session reads those files before it does anything, so it starts from where the last one left off.
An identity of its own. The agent has an email address. It sends and it receives. Something arrives, the agent picks it up, does the work, and mails the result back out. I can drop an idea in from anywhere by emailing it, the same way I’d email a colleague.
A way in from any device. SSH over a private network into the same tmux sessions. On the phone I use ShellDrop, an iOS terminal app I built for this exact job. From a phone on the train or a laptop in a hotel room you land in the same session, with the same half-finished piece of work sitting where you left it.
Building on request. Landing pages, dashboards, review pages, small internal apps, a report that runs every Monday. I describe what I want and watch it get built. The pages on this site were built this way.
A normal day through it
Most days start on the phone before I’m at a desk. Open a session, read what came back overnight, hand out the next few jobs. A client site needs a page, the ad account needs its search terms read, an email needs a first draft with the account history already in front of it.
Then I go and do the parts that need a person: calls, decisions, conversations with clients. The agents keep working on the machine at home. Work comes back through email or sits waiting in the session. I read it, correct what’s off, approve what’s right, and the next thing starts.
The part that surprised me is how much of the value sits in the memory. The model matters, but the files are what make the thing feel like it knows the business. Asking “what did we decide about this client last quarter” and getting a real answer with the reasoning attached takes a whole category of work out of the week.
What it replaced
It replaced re-explaining my business to an AI every single time I wanted something from it. It replaced notes scattered across five apps, none of which the AI could read. It replaced the version of my week where the ordinary work had to wait until I personally sat down to do it.
Practically, it’s what lets one person cover the marketing and the software for a business and still keep up. That’s the honest version of the claim.
Where it still needs me
It can’t see. It writes the CSS and has no idea whether the page looks right. Visual judgement is mine, every time.
It can be confidently wrong. Long jobs still get checked. Anything with client money or a client relationship attached gets read properly before it goes anywhere.
The memory only holds what gets written down. Skip the end-of-session write-up and the next session starts thinner. That discipline is on me, not on the agent.
It took time to build, and it needs maintenance. Files go stale, instructions need tightening, and something that worked last month needs a rethink when the tools move. Nobody hands you this finished.
And it needs someone to point it. The system has no opinion about what the business should do next. That part hasn’t moved an inch.
Why call it an operating system?
Because that’s the job it does. It organises what the agents can reach, holds memory between sessions, and handles what comes in and what goes out. The name is a plain description of the work. There’s no software in the middle of it beyond the agents themselves and the files they read.
What is an AI agent operating system?
An AI agent operating system is the structure you build around an AI agent so it can run real business work: a machine that stays on, an agent like Claude Code doing the execution, plain files that hold the business context and the history so nothing is re-explained, an identity such as an email address so work can arrive and leave without you, and remote access so you can drive it from any device. The agent is the engine. The operating system is everything around the engine that makes it useful on a Tuesday afternoon.
Do you need a framework to build one?
No. Mine is folders, Markdown files, terminal sessions and an email account. Every layer of software you add between yourself and the agent is another thing to learn, another thing to break, and another thing to migrate when the tools move. Files and folders survive all of that, and you can read them in any text editor without asking anyone’s permission.
Can you build this yourself?
Yes, and that’s exactly what the agent operating system course is: half a day on Zoom, two seats at most, starting from opening Terminal for the first time and finishing with your own version of this running on your own machine, loaded with your business, with your first custom app built on top of it. You leave with the system and the method, on your hardware, in folders you own.