What is an “Agentic OS” (and why you probably have half of one already)
The buzzword for using AI as a system instead of a chat. What it really means, what the pieces are and how I set it up at home.
A while ago I watched a TikTok that said something like: “Most people use Claude as a chat, and that’s why every conversation starts from zero. An Agentic OS is the opposite: your complete command centre.” The line stuck with me because it sums up a change I’ve been noticing for months in how I use artificial intelligence.
In this article I’ll explain what the term means, what the pieces are and how I have it set up, with real examples and no hype.
First, what it is NOT
“Agentic OS” sounds like a new operating system, as if you had to wipe your computer and install something other than Windows or Linux. That’s not it. The term is used in two ways and it helps to separate them:
- An operating system with agents built in. Microsoft, for instance, has talked about turning Windows into an “agentic OS” with built-in agents that do tasks for you. There are also research projects like AIOS that treat agents as if they were system processes. Interesting, but still early days.
- The way you organise your AI. This is what the video meant, and it’s something you can actually use today: building an environment with memory, tools and projects around an assistant like Claude, so it stops being a chat that forgets everything.
From here on I’m talking about the second one.
The problem with a chat that starts from zero
When you use AI as a normal chat, every conversation is a blank page. You explain who you are, what machines you have, what you tried last week… and next time, again. It’s like having a very smart colleague with amnesia.
An Agentic OS flips that around: context accumulates. My favourite line from the video was: “accumulated context is worth more than twenty benchmark points.” And it’s true: a slightly weaker model that already knows how your home network looks is more useful than a brilliant one you have to explain everything to.
The pieces
These are the pieces that usually show up in this kind of setup, with examples from mine.
1. Persistent memory
The agent keeps notes about you and your projects between sessions. Mine knows, for example, that my NAS changes IP address and has to be reached by name, or that my laptop runs the fish shell so SSH commands need to be sent differently. Things I used to repeat in every conversation.
2. One folder per project, with its own instructions
Each project has a folder and an instructions file (in Claude Code it’s called CLAUDE.md). In my study folder, for example, I ask it to act as a tutor: explain, let me try, then correct, instead of doing my homework. The same assistant behaves differently depending on where it’s working.
3. Tools and connectors
This is where MCP (Model Context Protocol) comes in: the standard for connecting the agent to outside things like email, calendar, documents or a browser. And above all, the terminal: being able to run commands is what turns a chat into something that does things.
4. Reusable skills
Procedures you repeat often, packaged as “skills”: how to create a document, how to review code, how to build a report. Instead of explaining the process every time, the agent loads the recipe.
5. Background and scheduled tasks
Agents that work while you do something else, and tasks that run on their own at a set time. You need to be careful here (more on permissions below), but used well it saves a lot of time.
6. The command centre
The visual part: a dashboard where you can see your projects, the state of your machines and your to-dos at a glance. I built one that runs on a Raspberry Pi and that the agent itself can update.
What it looks like day to day
A real example from this week. I asked the agent to set up Telegram alerts for when any service at home goes down. Because it already had context, it knew which services I run, where each one lives and which monitoring tool I use. It created the bot, configured the notifications, took a service down on purpose to check the alert arrived and then updated my dashboard to remove the “you have no alerts configured” warning.
With a normal chat that would have been twenty messages of copy and paste. This way it was one conversation.
What nobody tells you: permissions
An agent that can run commands on your machines and log into your accounts is powerful, but also sensitive. Two things I’ve learned:
- Let it ask too often. I’d rather it asked before deleting something or touching a machine I’m using. Once, during some tests, it overloaded my laptop while I was working on it. Since then, heavy tests get announced first.
- Keep it away from credentials unless needed. My agent’s permission system once blocked it when it tried to look for an old token inside a backup. At first it felt over the top. Then I realised it’s exactly what I want: passwords are better in my hands.
Golden rule: give the agent just enough power for the task, and widen it once you trust how it uses it.
Is it worth setting up?
If you use AI now and then for one-off questions, no. A chat is plenty. If you use it daily for work, study or tinkering, yes: the productivity jump doesn’t come from a better model, but from not starting from zero every time.
My advice is to go one piece at a time: first a folder per project with its instructions, then memory, and only at the end connectors and automated tasks. And the dashboard, once you have something to look at on it.
Did it stick?
Three quick questions. Each right answer is worth 10 XP.