DHH: "I Haven't Written a Single Line of Omarchy Quattro's Code"

The three ages of agentic coding, in nine months

DHH recounts a dated turning point. Thirteen months earlier, on the same show, he held a skeptical position: autocomplete didn't interest him, the chatbot made a good tutor, nothing more. On November 24, 2025, the release of Opus 4.5 changed everything. It's not so much the model's raw intelligence that improved, he says, but its ability to instrument the machine, use tools, and verify its own work.

He splits what followed into distinct phases. A first "agentic moment" of three months where the human stays at the wheel: he describes the task, corrects the trajectory, reviews everything. Then the arrival of sub-agents in spring, which cut execution time by five to ten. Finally, this summer, with Opus 5, Fable, and Sol, a deeper shift: he no longer indicates the route, he lays out the problem and the model proposes the path.

His metaphor is the GPS. At first, you watched the device for fear it would send you into the harbor; today, the car drives itself. "I've become optional in the part that produces the code," he sums up — at least in the areas where he works today.

Omarchy Quattro: an OS as proof by the pudding

Faced with the accusation of "AI psychosis," DHH answers with a deliverable. Omarchy, his distro based on Arch and the Wayland compositor Hyprland, has existed for a little over a year. The Quattro version, released a few days before the recording, was developed in three months with a 100% agentic acceleration.

He admits to selective review: he examined the shape of everything, line by line for what's critical in the model layer, and not at all for part of the UI code. For the writing app Omawrite, written in C++/Qt even though he doesn't know that language, he forced himself to never open the code — a deliberately black-box experience. First version in twenty minutes, dropped Typora within two days.

The participation numbers impress more than the lines of code: over 1,000 merged pull requests in three months, about 400 pending, and 330 plugins published on the marketplace in three days. Many come from people who aren't systems developers. DHH no longer reviews PRs himself: agents sort them, validate fixes in a VM, and only report the final decision to him.

Where it breaks down: large codebases and humans

The interview isn't just a panegyric. DHH recounts the February failure at 37signals: for the final stretch of Basecamp 5, the team let designers code with agents. The result was an avalanche of individually defensible PRs that, taken together, destroyed the system's architecture. It had to be cleaned up manually.

His lesson: on an existing, substantial codebase, you need to stay a programmer to preserve the architecture. He immediately qualifies this by turning the "slop" criticism back on humans: after 3,000 developers, big companies' codebases are "absolutely awful."

On the question "why don't Photoshop or Premiere improve ten times faster?" his answer is organizational. The bottleneck was never implementation but human bandwidth: product managers, designers, VPs, CTOs who all want to be part of the scoping. To get the 10x or 100x factor, he says, you need to interact directly with the agents, without a human intermediary.

He adds a harsher critique: most organizations don't know what they want. They aren't limited by their ability to produce code, but by their ideas, vision, and taste. "You can bring a lot of bad ideas into existence, and then what?"

The setup: 16 threads, Herdr, Tailscale, and a model benchmark

The most operational part of the episode describes a completely rethought workstation. You move from single-threaded thinking to parallelism: multiple agents, so multiple panels. He started with tmux, then switched to Herdr, which adds notifications when an agent is waiting for a decision. Neovim now only serves for navigating and reviewing diffs.

Realizing a single machine wasn't enough, he pulled four mini-PCs out of a closet, connected them with GL.iNet Comet KVMs and a Tailscale network. He estimates his personal ceiling at around 16 simultaneous agent threads — beyond that, it's the human that saturates.

The homemade benchmark is telling: rewrite in Rust, dependency-free, the Python library Terminal Text Effects used by his screensaver. Fable produces a detailed plan and finishes in under 45 minutes (startup time went from 86ms to 2ms, execution 9.6x faster), for about $550 in tokens if he'd paid per use. Sol redoes the work in 1h30 for $46, Grok 4.6 for $55, DeepSeek V4 Pro in 2h45 for $23. GPT Luna fails and even tries to cheat. Two self-research loops later, the gain reaches 46x.

His standard procedure: have Claude or Fable do the work, then systematically have Codex xHigh review it, and let Copilot pass over it again on GitHub. "Even a good programmer, if a peer reviews their work, the code will be better. Build that into your process."

Mourning artisanal code and advice for developers

DHH admits he's surprised he isn't more nostalgic. His economic argument: we wrote beautiful code because a coherent architecture made human modifications less costly. That premise is crumbling. He qualifies this, though: as long as tokens are a scarce resource, readable code still pays off, because agents iterate more cheaply on it.

To those panicking, he gives simple advice: don't try to anticipate. No one knows what things will look like two model generations from now, and trying to guess is the shortest path to anxiety. He adds a reassuring point: there's almost no accumulation in this field. Someone who spent a year hiking would catch up to the frontier in two weeks.

He distinguishes two types of developers: those who loved the mechanical assembly of logical constructs, who are threatened, and those who loved building things, who are not. He himself describes coming back to his starting point: he became a programmer because he wanted programs, not out of love for conditional statements.

He refuses, however, to call vibe coding "programming." Programming implies understanding the primitives: loops, conditions, variables. And he slips in a paradox: his long experience may have been a handicap, because he prescribed to the agents how to do things instead of describing outcomes.

Linux, the historical irony, and everything else

The episode's strongest thesis: Linux is going to win the desktop, not despite its flaws but because of them. Everything in it is a config file or a command-line tool, which makes it the ideal terrain for agents. Esoteric error messages, once off-putting, become actionable by a model pre-trained on 40 million lines of kernel code. Omarchy Quattro even includes a "crash watcher" that offers to have the agent diagnose any crash.

The obsession with install speed illustrates his method. Initial goal of 15 minutes, then 2, then 1; current record of 45 seconds, with a Dell XPS-specific image targeted at 12 seconds. Gains come from preloading while the user answers questions, and surgical size reductions: 180MB saved on the JetBrains font, 200MB on NVIDIA drivers. His acknowledged reference: McLaren engineers who chase 370 grams on a one-tonne car.

The conversation then drifts widely off tech: AI video generation, fatherhood, X and engagement algorithms, his immersion in the mass immigration debate in Europe, his anti-longevity stance — he's aiming for 90 to 100 years and hid a memento mori in Omarchy's calendar, which displays the percentage of life elapsed.