Linux Setup Was a Hazing Ritual. Agents Changed It.

TL;DR: Developers didn’t pick macOS because Linux was a worse operating system. They picked it because Linux setup was a skills test, and the test was reading wikis until you found the one command that fit your hardware. An agent turns that test into a conversation. I hit six breakages across two Linux machines, and by my own estimate most took five minutes or less with an agent, where each would have cost two to five hours before agents. Two didn’t: the keyboard and the scanner. The agent doesn’t remove the work. It removes the research, and that was always the hazing part.

Why do so many developers still pick macOS?

Someone on X asked why so many of us still pick macOS. My quote-post answer: “Because the setup of Linux has been treated like a hazing ritual of config and driver hell for decades?”

Ask a macOS developer why they stayed and you’ll get “it just works.” What that usually means is the first week, which on Linux was where you earned your seat.

What was the old loop?

You hit a symptom. You searched for it. You found a forum thread from 2019 and a wiki page written for a different distro. You tried one command, then the next, because changing two things at once meant you couldn’t tell which one helped. Often you first had to coax the system into producing an actual error message, because the failure was silent and the real question was which component was broken.

Then, once you knew what was wrong, you worked out the config change. My estimate for these problems, before agents, is two to five hours each.

That loop is the hazing. What it measured was patience.

What actually broke on my machines?

I run two Linux machines now: an M1 MacBook Pro on Asahi that dual-boots macOS, and an x86_64 desktop that dual-boots Windows 11. Six things broke, in this order:

1. Aquamarine, mid-install on the M1. The Omarchy installer tried to pull a version of aquamarine, the rendering library Hyprland depends on, that was too new to install. The machine was busy installing, so I used Claude on my phone. It did the research and wrote out the commands, and I applied them myself. The fix was a downgrade to an older build from the pacman cache. I wrote that up in the M1 field report.
2. World of Warcraft retail. Mouse input wasn’t recognized. I couldn’t scroll to accept the terms of service, and some clicks were ignored. It turned out to be how Hyprland translated input into the game’s window. That was on the x86 desktop, fixed with Claude Code, and relatively quickly.
3. Keyboard function keys and some other inputs. These were broken from the start on the desktop, and I debugged them with the agent. This one took a little while: pressing keys, testing back and forth, checking which signal actually reached the machine.
4. A trading-card scanner. It didn’t work out of the box on the desktop. With Claude Code, getting it to scan didn’t take long. Getting it good enough to actually work with took longer: scan after scan, tuning saturation and color until the images were right. On Windows that tuning was easy in the vendor’s PaperStream software.
5. A thermal printer. It wouldn’t print at the right size and wasn’t visible on the network.
6. A Canon PIXMA printer. Also invisible on the network. I debugged both printers with the agent.

None of it is exotic. It’s the ordinary tax of an OS that hardware vendors don’t write for first.

What changed, and what didn’t?

By my own estimate, not a measurement, most of these were solved within five minutes or less of using the agent. The keyboard and the scanner took longer, and I won’t round them down to match the others.

The fair objection is that I just got better at Linux. Maybe partly. But the hours are my estimate for these same problems without an agent.

What changed is who does the research. The agent reads the forum threads and the wikis, then turns them into commands for the thing I actually described. I still verify that the result is the result I wanted. No agent can tell me a scan’s saturation looks right, and none presses the keys on my keyboard.

What didn’t change is that the human owns the outcome. An agent is a fast colleague who has read everything and never seen your desk. You still read what it did and check the result.

Is it safe to let an agent fix your OS?

Not by default. Start it from Omarchy’s launcher and the agent fixing your system likely runs with approvals skipped. I went through that launcher in the agent safety post: eleven of the fourteen agents it knows how to start get launched with their own spelling of “don’t stop to ask.”

Omarchy’s own manual labels its config skill experimental, recommends plan mode first, so you read what it intends to do before it does it, and names omarchy reinstall configs as the way back if the agent makes a mess (quoted in the agent safety post). That skill covers config files, though. A printer or scanner driver needs root, and Omarchy’s own rule, quoted in the AGENTS.md read, routes privileged commands an agent launches through a password prompt.

For a printer, I’m comfortable answering that prompt. For anything that touches disk layout, a bootloader or credentials, I want the plan in front of me first. Same agent, different blast radius.

Am I the only one asking an agent to change the OS?

No. Tobi Lütke, CEO of Shopify, said on The Knowledge Project that he changes his operating system by asking an agent and never touches the config files. That’s my paraphrase of the segment starting around 21:51, from an episode uploaded September 15, 2026. It’s the same shift from a very different seat.

Why is the macOS answer shifting?

macOS earned its place with developers by being the machine where you didn’t pay the setup tax. That was a fair reason for a long time.

The reason is weaker now because the tax is smaller. When the cost of a driver problem drops from an afternoon to a conversation, “it just works” stops being a strong enough argument for a developer who’s comfortable in a terminal and willing to verify what an agent hands back.

It’s true on my two machines, with my tolerance for fiddling, and with a recoverable system underneath. It isn’t true for a laptop you need on Monday morning, a fleet of company machines, or anyone who doesn’t want to be the person who verifies.

But the hazing was never the point of Linux. It was a side effect of nobody being around to answer your question. Now something is.

Read this next: my field report on the M1 install, or the rest of the Omarchy hub.