Every agent I hand a new machine to starts the same way. It runs ls. It opens a config file that moved two releases ago. It picks the wrong package manager, gets corrected, and apologises. Ten minutes later it knows things I could have told it in one line.
For a long time I read that as the model being careless. It isn't. Nobody had told it anything.
We hand an agent a terminal and call that giving it access. It works the rest out by touching things. Which init system this is, where the compositor keeps its config, whether that service is masked or just stopped. It cannot see that something crashed four seconds ago, because nothing routes that anywhere near it. It has a shell and no standing. That works the way handing someone a keyboard and no password works, and you pay for it again in the next session, and the one after that.
I have been running Omarchy for a while now and I like it more than I expected to. It started as a good Arch install that somebody else had already argued about, which was enough. Quattro shipped last week and it turned into something I actually want to write about.
The interesting thing in it is not the agent picker. It is that the release quietly contains two manuals. One is written for me, a proper guide sitting in the menu, and DHH is blunt about why it exists. He thinks software should ship with good manuals. Fine, and rare, but not new.
The other one is written for my agent. The system installs skills that describe itself, including one that teaches an agent how to reconfigure the machine and one for reading a crash. His own summary of the point is better than anything I would write. Ask your agent to do something on a fresh install and it is going to know stuff, and the reason it is going to know stuff is that the system shipped skills.
That is smaller and more boring than the phrase agentic operating system suggests, and I think it is the whole thing. The machine writes documentation for a reader who is not you.
Once you look at it that way the rest of the release stops reading like a feature list. The agent gets picked during setup next to the keyboard layout and the timezone, and installs itself the first time you reach for it. When something segfaults you get a notification, and clicking it hands the core dump over with a skill for reading backtraces, and it knows how to write the bug report properly if the problem turns out to be Omarchy itself. The agents run on mise instead of system packages, which sounds like plumbing and isn't, because they ship several times a day and nobody wants to update their whole machine to get a fix. Every one of those is the same move. Stop making the agent infer something the system already knows.
Knowing what to do is not the same as it being safe to do, and this is where most talk about agentic desktops falls over. Omarchy is a system package now rather than a git checkout in your home directory, which is how it used to work and how it used to break. Your changes sit on top as their own layer. You can clone a built in plugin and have your copy win, and an update will not reach in and overwrite you. Every update takes a snapshot first, and you can boot the one from before. Let the agent change the system, and make every change cheap to undo. Neither half is worth much alone. A machine an agent cannot modify is a machine it cannot help you with, and a machine you cannot roll back is one you will never let it near.
The pitch underneath all of this is not really about agents at all. DHH has said he put thousands of hours into his own setup, and almost nobody else is going to do that, which is why most people run an operating system they are not allowed to have opinions about. His demo was pointing an agent at a web page for a Mac app he liked and asking for something similar as a panel for his bar. He got one. He did not pick the details and cheerfully said so. Malleability used to be gated on time. If it stops being gated on time, an opinionated system with good defaults and a documented way in is worth more than a system that is merely configurable.
The browser is the same argument one layer up and it is further behind. Agents drive browsers today by pretending to be a mouse, through coordinates that break on a layout shift or selectors that break on a refactor. Meanwhile the accessibility tree sits right there, already maintained, already describing what everything is and does.
So my guess is the division that ends up mattering is not agentic against normal. It is systems that describe themselves against systems that make the agent work it out. Inference is where the tokens go and where the mistakes come from.
None of which makes it the right answer everywhere. An agent that is a real user of your machine has your blast radius, and that is the point of it and also the problem with it. For anything I would not hand my own login to I still want a container, and the plugins are not sandboxed, which DHH says plainly rather than papering over. This is for the machine you already trust the agent on. It is not a security boundary and it does not claim to be.
One last thing, because it made me laugh when I found it. The Quattro screensaver picks a random effect from terminaltexteffects, which the Omarchy folks rewrote in Rust as ttfx. I sent a couple of small fixes that way, so I went and read how one of them actually works.
A screensaver is a strange place to find the thesis of a release. But it is the same move as the skills. Somebody decided the machine should describe itself well enough that whatever is reading it does not have to guess.