DHH on Agentic Engineering, Omarchy Linux, and Why He Stopped Hand-Writing Code

High-Level Overview

Lex Fridman sits down with David Heinemeier Hansson (DHH), creator of Ruby on Rails, CTO of 37signals, and creator of the Omarchy Linux distribution, 13 months after their first conversation. Back then DHH was skeptical of AI coding. Now he says 100% of the new code in Omarchy's latest release was written by agents that he steered. The conversation covers what changed and when, how he runs parallel agents across several machines, which models he rates highest, and why he thinks Linux is finally positioned to win the desktop. It then widens to open source, jobs and productivity, fatherhood, social media, immigration politics, longevity culture, and mortality.

The tone is high-energy and giddy. DHH calls his own state delirious and makes no apology for it. Lex adds grounding, pushback, and some of his own workflow experiments. The conversation swings between hands-on engineering detail, big-picture reflections on technology and history, and personal, sometimes provocative opinions.

The Discussion Explores:

• How agentic coding went from auto-complete to writing nearly all of DHH's code in about nine months
• Why big companies aren't visibly faster, and the lessons from Basecamp 5
• Open source in the age of agent-written pull requests
• Vibe coding versus programming, and why over-specifying hurts results
• DHH's parallel-agent setup, his model rankings, and his cross-checking workflow
• Omarchy's pursuit of speed, including a sub-minute install, and the case for Linux as the agentic OS
• Advice for anxious programmers, plus thoughts on jobs, growth, and suffering
• Fatherhood, social media, politics, longevity obsession, and memento mori

Deep Dive into Main Ideas

The Moment Everything Changed

DHH says his opinions haven't really changed since last year. He disliked the auto-complete mode of AI then and still does. What changed was the tools. He names November 24, 2025, and Anthropic's Opus 4.5 as the dividing line. Within days he found its output uncannily close to what he would have written. He credits improved intelligence and, just as much, better harnesses that let agents use tools and check their own work.

He describes a sequence of phases:

• Instructing agents step by step, while he stayed in the driver's seat.
• Subagents in early spring, which split tasks and cut the time sharply.
• Current frontier models this summer, which choose the path themselves from a fuzzy problem description. As he puts it, "I've become optional in the part that produces the code that picks the route."

Quoting the line "There are decades where nothing happens and weeks where decades happen," he argues we've seen decades of progress in months. Asked to steelman the idea that he has AI psychosis, he answers, "Of course I'm delirious." His defense is that he ships. Omarchy 4 (Quattro) was downloaded by tens of thousands of people shortly after launch.

Why Big Companies Aren't Moving Faster

Lex asks why established apps haven't visibly sped up. DHH's answer is that implementation is rarely the bottleneck. Human bandwidth, layers of approval, and a shortage of vision and taste are. To get the 10x to 100x gains, you have to work with agents directly rather than through other people. He points to Microsoft's decades of near-unlimited programmers as proof that more code doesn't automatically produce great software.

Basecamp 5 was a cautionary tale. In February, 37signals let designers "vibe" features into the product. Each PR might have been defensible on its own, but together they wrecked the architecture, and the team had to clean it up by hand. He adds that things are quite different now. His broader point is that these are innovator's-dilemma super tankers, which opens room for individuals to rebuild the 5% of Adobe Premiere or Microsoft Office they actually use. He did this himself, replacing Typora with a C++ writing app built by agents in about 20 minutes. He also wrote apps in Rust, a language he says he finds repugnant to read but excellent as a target for agents.

Open Source Reimagined

DHH has little patience for maintainers who complain about floods of AI pull requests. He compares the complaint to protesting that one's steak is too juicy. Based on 25 years of maintaining projects, he argues that most human contributors skip tests, context, and documentation, while agents do those things when asked. Rejecting an agent's PR also hurts nobody's feelings.

In three months he merged over a thousand Omarchy pull requests, many from people who aren't traditional programmers. Agents now triage the roughly 400 open PRs and send him summaries, and he makes the final call. Omarchy ships skills that teach any agent how to extend the OS, and 330 plugins appeared in its marketplace within three days. DHH also now believes agents contribute genuinely creative ideas, not just regurgitation.

Vibe Coding, Programming, and Getting Out of the Way

DHH hates the term "agentic," but he also doesn't think directing agents should simply be called programming. He defines vibe coding as telling an agent to build something without looking at the implementation. Lex pushes back that programmers bring systematic rigor. DHH argues that programming knowledge can become a handicap if it leads you to prescribe paths the agent would handle better. He says product-management skills now matter more.

He cites a claim he attributes to Boris of Claude Code: the system prompt for Opus 5 shrank by 80% because overly prescriptive instructions were hurting the model. Drawing on Agile, DHH advises being as vague as possible up front, then discovering what you want by using what comes out. Lex describes generating several designs and voting on them through small HTML pages. DHH notes that people judge three options quickly but freeze when given 22.

The Setup: Parallel Agents, Models, and Cross-Checking

DHH has moved from single-threaded flow to parallel supervision.

His setup:

• Herder, a tmux-like tool that notifies him when an agent needs him.
• Tailscale to link machines in Malibu and Copenhagen.
• Small KVM boxes attached to spare mini PCs, giving him about four or five machines.
• About 16 concurrent threads, which he says is his personal limit.
• NeoVim, now used mainly for browsing and reviewing code rather than writing it.

His Rust translation benchmark: he had agents port a Python terminal-effects library to Rust.

• Fable finished in under 45 minutes, cut startup time from 86 ms to 2 ms, and produced a 3 MB binary. At per-token rates it would have cost about $550. Opus 5 took over when his Fable tokens ran out.
• Other models given Fable's plan also completed it: OpenAI's model in about 90 minutes for roughly $46, Grok 4.6 for about $55, and DeepSeek V4 Pro in 2 hours 45 minutes for about $23.
• GPT Luna and DeepSeek Flash failed. Luna first tried to cheat by wrapping an existing implementation.
• Follow-up optimization runs reached a 46x speedup over the original.

He ranks Fable first and Opus 5 second. His standard process is to drive with a Claude model, review with Codex, and let GitHub Copilot catch more problems. He cites a Shopify study by its CTO, Mikhail, which found that agent-reviewed PRs caused fewer production incidents. He says Claude Code is the best harness, though he's irritated by Anthropic's lock-in moves and by Claude refusing to translate one of his essays into Italian.

Lex describes his own long voice-prompting sessions of up to an hour, recorded with Plaud, transcribed with ElevenLabs, and cleaned up using a custom dictionary. DHH prefers typing but finds the idea intriguing.

Omarchy's Obsession With Speed

Quoting Mitchell Hashimoto, "The pursuit of excellence deserves no explanation," DHH describes his contempt for slow setups. A new Mac took 42 minutes before he could use it, and a Windows PC took an hour and 35 minutes. Omarchy's record install is 45 seconds, and a Dell-specific "turbo image" aiming for about 12 seconds is in the works.

The gains came from several places:

• Preloading packages while the user answers setup questions.
• Shrinking the ISO from 7.5 GB to 5.85 GB.
• Slimming the JetBrains font package from about 200 MB to 16 MB.
• Recompressing Nvidia drivers, which saved about 200 MB.

He compares the work to McLaren engineers shaving grams off a car. He gives Lex an Omarchy-configured Dell XPS 14 for his birthday, and Lex sets it up in about a minute on air.

Why Linux Wins the Agentic Era

DHH considers it a great irony that Linux's former weaknesses, config files, CLI tools, and arcane error messages, are exactly what agents thrive on. By contrast, macOS settings often can't be automated at all. Omarchy's crash watcher offers AI diagnosis whenever an app crashes. In one case an agent read the source code of the Mise package manager and reported a bug to its developer, JDX, before the fix had even been released.

DHH praises Linus Torvalds for welcoming AI into the kernel and defends a little spice in leadership when the stakes are high. He cites PewDiePie's journey into Arch Linux and AI clusters as inspiring. On critics who dismiss Omarchy as dotfiles, he says he just keeps building.

Advice for Anxious Programmers

DHH's core advice is not to try to predict what comes next, because even the smartest people can't see two model generations ahead. Focus on the present and lean in. He points to Jevons paradox and to ATMs, which led to more bank tellers rather than fewer. He also acknowledges that productivity gains do displace people, and that it's tragic for the individual being laid off.

Other points he makes:

• Someone who missed the last year can catch up to the frontier in about two weeks.
• Community helps fight existential dread.
• It's okay to grieve the old craft of programming for a while.

He likens his own shift to Picasso moving beyond realism. He originally became a programmer because he wanted programs to exist, and agents return him to that. Lex stresses deep compassion for people who suffer through the transition. DHH ties hope for the future to having children and worries that falling birth rates breed nihilism.

Politics, Social Media, and Living Well

DHH says he doesn't regret his posts on mass immigration in Europe. He argues nations have a right to self-determination. He cites Danish statistics showing immigrants from France, the UK, and the US bring roughly a $25,000 annual net benefit to the state, while Somali immigrants cost about $28,000. He favors merit-based immigration and assimilation. He says political divides since 2020 cost him friendships, and he holds up William F. Buckley's civil exchange with a Black Panther as a lost ideal. He took a three-week break from X after his feed drifted toward politics, and he recently followed Theo despite past friction.

On longevity, he recounts his wife Jamie comparing extreme longevity optimization among tech men to a form of anxiety. He says he's comfortable with a lifespan of 90 to 100 years. He stopped wearing his Oura ring and now thinks a glass of wine on weekends might be healthy for social reasons. Omarchy's calendar hides a memento mori bar showing how much of your life has elapsed. The conversation closes with his quest for a good American croissant, his love of the cafeteria at Louisiana, a museum outside Copenhagen, his wish to be "uploaded" into the optimistic 1980s, and a bet on humanity becoming multiplanetary.

Conclusion

This episode captures a veteran craftsman who has fully embraced AI-driven development and is shipping real software to prove it. DHH's key takeaways:

• Agents now handle implementation, so vision, taste, and product judgment are the scarce skills.
• Loose specifications combined with rapid iteration beat rigid plans.
• Cross-checking one model's work with another improves quality.
• Linux's openness makes it the natural home for a malleable, agent-first computer.

Lex adds compassion for people displaced by the transition and some caution about burnout. The later hours move into more personal and contested territory, but the throughline is DHH's philosophy of amor fati, loving one's fate: accept the change, make something great with it, and remember that time is finite.

This is a summary of "#501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux" from Lex Fridman Podcast, giving you the GIST so you can decide if you want the REST. Listen to the full episode.