DHH on Omachi Linux and AI Coding Agents

For twenty five years David Heinemeier Hansson built his name on a single, almost monastic skill: writing Ruby code clean enough to read like prose. He is the creator of Ruby on Rails, the framework behind Basecamp, Shopify, and thousands of smaller web businesses. So it is worth pausing on what he says now. In the last two months of building Omachi Linux, his own operating system, he has not written a single line of the shipped code by hand.

"I have not written any of the code that shipped in Quatro by hand. I've reviewed the shape of all of it... but I've not written any of it entirely by hand." — David Heinemeier Hansson [21:41]

Quatro is the newest release of Omachi, a Linux distribution Hansson started building last summer between racing sessions at Le Mans. He calls the last three months of its development the clearest evidence yet that something changed in how software gets made. That change, in his telling, has a precise start date: November 24, 2025, when Anthropic released a model called Opus 4.5. He did not even try it on release day. Two days later he gave it a few tasks and the output looked uncannily like something he would have written himself (13:22). "There are decades where nothing happens and weeks where decades happen," he says, borrowing the line from Lenin to describe what followed (10:03).

It is easy to hear that kind of talk as hype. Hansson knows this, and he admits he was, until recently, the skeptic in the room. Shopify's chief executive Tobi Lutke had been telling him for months that something big was happening. "I was a skeptic... I was still on Earth. And he had already boarded the rocket and was heading towards space," Hansson says of that period (51:24). What convinces him now is not a chart or a benchmark. It is a specific, almost domestic story: he needed a lightweight writing app to replace a piece of shareware called Typora, told an AI agent to build it in C++ and Qt, and had a working version in about twenty minutes. Two days later he had abandoned Typora for good, and he has written every essay since inside the tool the agent built, a program he calls OmaWrite and has never personally inspected line by line (32:17).

What the machine is actually bottlenecked on

The more interesting argument in the conversation is not that AI can write code. Plenty of tools have claimed that for years. It is Hansson's claim about where the real bottleneck in software has always been: not implementation, but communication between humans. "To get that magical 10X, 100X, in a few rare cases 1,000X productivity boost, you have to interact with the agents directly, and you cannot intermediate that bandwidth with another human because it's simply too slow," he says (25:19). Large companies, he argues, are not slow because writing code is hard. They are slow because a product manager, two designers, a vice president, and a chief technology officer all have to agree on what to build before anyone builds it.

This is not an abstract claim for him. Basecamp's own team learned it the hard way earlier this year. When designers were given agent access and told to "vibe" their way through building Basecamp 5, the individual pull requests looked fine on their own but collectively wrecked the underlying architecture, and engineers had to clean it up by hand (22:41). The lesson, as he tells it, is not that non-programmers should stay away from agents. It is that judgment about system shape still matters, even when the typing is automated.

That same logic extends to open source, where Hansson has maintained projects for about 25 years. He has merged more than 1,000 pull requests into Omachi in the last three months alone (40:02), a volume that would have been unthinkable for a project of its size a year ago. Many came from people who were not traditional programmers but had ideas an agent could translate into working code. His view of maintainers who resent this flood of contributions is unsentimental: "I would rather get an agent-written pull request to one of my projects than I'd get one written by a human, and it's not just because the quality's better," he says, adding that rejecting an AI's work costs no human feelings (38:41).

Speed as a philosophy, not a metric

Some of the most vivid material in the conversation has nothing to do with intelligence and everything to do with obsession. Hansson has been chasing the installation time for Omachi down from minutes to seconds, and the current world record for a full install stands at 45 seconds (126:07), against the 42 minutes it took him to set up and update a brand new Mac before he could even open an app (120:08). He justifies the pursuit with a line borrowed from Ghostty terminal creator Mitchell Hashimoto: "The pursuit of excellence deserves no explanation" (118:48). The point is not that a faster install matters to most users. It is that the search for speed forces discoveries, like preloading packages while a user answers setup questions, that make the whole system better.

Underneath the showmanship is a real economic calculation about which AI model to use for which job. Hansson ran the same test, translating a Python animation library into a dependency free Rust program, across several models. Claude's Fable model did it in under 45 minutes and produced the cleanest plan, but would have cost about $550 in token spending outside a subscription (160:26). DeepSeek's Pro model did the identical job for about $23, taking roughly two hours and 45 minutes (163:02). His conclusion is practical rather than ideological: use Fable to plan and to review, and use Opus 5 for the grinding work of implementation (164:14).

He extends the same reasoning to hardware itself. Linux has spent three decades losing the desktop war to Windows and macOS partly because its strength, file based configuration and command line tools, was also its weakness for ordinary users. Agents flip that. A large language model trained on millions of lines of Linux source code can read an arcane kernel error message and fix it faster than a human searching forums ever could (150:35), which is why Hansson insists Linux, not Mac, is now "the ideal platform for AI agents" (107:51).

None of this erases the human cost of the shift, and Hansson does not pretend it does. He argues, somewhat bluntly, that productivity gains have always meant fewer people doing a given job, and that this has always been the price of economic growth, citing the historical shift away from field labor as an example (80:26). Late in the conversation, discussing the extreme longevity movement popularized by figures like Bryan Johnson, he repeats an observation from his wife that struck a nerve online: that the pursuit of radical life extension in some men resembles anorexia in some women, a physical expression of anxiety and a wish for control (300:15). It is a strange but fitting closing note for a conversation about a man who spent 25 years perfecting one craft and then, without much mourning, handed most of it to a machine. He built a small memorial to that fact into Omachi itself: a hidden feature on the calendar that, once you enter your birth year, quietly tells you what percentage of your expected lifespan has already passed.