David Heinemeier Hansson
Related people
Episodes
- @DHH on startups and parenting in the new age
- David Heinemeier Hansson
- #501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux
Notes
Lex Fridman Podcast · #501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe…
-
Why Does Switching Between AI Agents Feel Like Racing?
David Heinemeier Hansson finds the switching mentally exhausting in an exhilarating way. It matches the feeling of stumbling out of a race car, barely able to hold his head up, and being glad to be alive. Racing tiredness is mostly physical. Rain adds a mental load because the car stays at the knife's edge. Le Mans gives time to relax on a long straight where the car is simply going fast in a line. Other tracks come at you the whole time. An hour in each leaves a very different kind of exhaustion.
Running agents at the limit of what a person can follow is the no-rest version. There is no passive stretch and no coasting: a problem, then the next problem, constantly in the corner. He thinks t
-
Why Does Basecamp Treat Coding Agents Like Coworkers?
David Heinemeier Hansson says Basecamp has started putting coding agents inside the product and treating them as coworkers. An agent gets a task on a to-do or on a card and does the work there. A collaboration tool built for asynchronous communication fits that better than a chat-style harness. Chat was the original shape of these tools, and it invites you to sit and wait for a reply. A Basecamp to-do does not. You hand it over and you do not expect to hear back immediately.
He is also building an Omabot that does autonomous development on Omarchy on a schedule, working through to-dos, pull requests, and issues, then emails him with the Hey command-line tool. One of those emails listed 12 p
-
Why Are Linux Error Messages Useful to AI Agents?
David Heinemeier Hansson says agents have taken the hardship out of diagnosing Linux and turned its overly specific, arcane error messages into an advantage. A message that makes no sense to a person is something an agent can place, especially when it was pretrained on 40 million lines of Linux code and knows where to look. He has not had a single problem on his Linux machine since the beginning of this year that an agent could not diagnose. A year and a half ago he was still searching forums for esoteric answers. Now the agent can see the source of the system and of the other software on the box.
Omarchy Quattro added a crash watcher just before it shipped. If any app crashes, a prompt ask
-
Why Does Omarchy Ship With Apps Already Installed?
David Heinemeier Hansson ships Omarchy with software already installed because the name is a chef's menu. The Oma in Omarchy is short for omakase, which means chef's choice. He makes the choices, and the install is the set of applications he thinks is awesome. In parts of the Linux world it is not acceptable for a distribution to decide which programs people use. Omarchy decides.
For video it ships with OBS already installed for recording, and with Kdenlive, the timeline editor he uses for his own videos. He also included Omocut, a clip cutter he built so a podcast clip does not require a full timeline editor. Control-Space sets the start, Alt-Space sets the end, and Control-S saves. NeoVim
-
Why Did Omarchy Shrink Its Install Image?
David Heinemeier Hansson cut the Omarchy install image from 7.5 gigabytes to about 5.85 gigabytes. Most of the install time is decompressing the compressed packages on the image, so the smaller size showed up almost one-to-one in how long setup takes.
The JetBrains font package on Arch is 200 megabytes because it includes every variation. Omarchy uses the monospace version in one nerd-font patched edition, about 16 megabytes, so a slim package saved 180 megabytes. He did the same kind of trim to the NVIDIA drivers. Coming straight off the Arch repo, they were not using the strongest compression, ZSTD, which is slow to create and worth it when the image gets smaller.
He compares the obsessi
-
Why Do Coding Agents Work Better on Linux?
David Heinemeier Hansson says coding agents work better on Linux than on Mac or Windows. They like the Unix idea of separate tools called from the command line, and on Linux almost everything is a config file or a command-line tool. Those used to be the reasons people avoided Linux. They are now the selling points, because an agent can read the files, change them, and copy a setup from one machine to another.
A Mac still hides pieces of the setup. Homebrew has gotten quite good as the missing package manager, but Raycast has no config file you can simply open. The export goes onto a USB stick and gets imported somewhere else. Default Mac key bindings are a manual click-through. Linux inside
-
Can a Programmer Catch Up on AI Coding in Two Weeks?
David Heinemeier Hansson says a programmer who missed the past year can catch up to the frontier in two weeks. There is no long accumulation you have to replay. Someone who spent a year hiking the Himalayas could come back and catch up in the same two weeks. Leave in October and return in April or May, and the change can feel like a century in a cryo chamber: what do you mean we are not programming the old way anymore?
The requirement is a willingness to become a different kind of practitioner, because the work is moving that fast. He also says grief is a fair response. It is not his own disposition, but a sense of loss for a world you knew and liked is a time-tested way to cope. He feels s
-
What Should Young Programmers Do About AI?
David Heinemeier Hansson tells a young programmer not to anticipate where the tools go next. Even the smartest people in the business cannot see what two model hops from here will look like, and trying to deduce two years out is a waste of time that can produce what he calls an AI psychosis. Focus on right now. Right now is the most incredible time to be into computers, and leaning in is what makes them do amazing things. Learn where the state of the art actually is, at least for a hot moment.
Building in public is optional. Open source is still worth joining because the camaraderie is enormous, and other people can pass on the useful kind of excitement: the future looks bright, and buildin
-
Why Does Cheaper Software Raise Demand for Programs?
David Heinemeier Hansson says programmers who love building things are not under threat from AI, and there is a case for needing more builders than exist now. Coping gets hard when the only love of programming was the mechanical work of assembling logic so a program does what someone else already asked for. That mechanical process is under threat. People who studied computer science, dreamed of building, and counted on a high salary can feel a real anxiety, and his own optimism is not evenly shared.
Employment numbers are fuzzy. Some stats show more openings because AI is dramatically lowering the price of programs, so people want far more of them. That is the Jevons paradox: when the price
-
Why Do Programmers Romanticize Writing Code by Hand?
David Heinemeier Hansson says a Commodore 64 programmer dropped into 2026 would be lost for a little while. That machine had a one-megahertz CPU and 64K of memory, and everyone who wrote software for it internalized constraints, because the machine was nothing but constraints. Efficient software is still beautiful. The optimization required to get a video game out of one megahertz is a different job from the optimization required now. Handwritten programming, slow and tightly constrained, felt like chiseling.
He treats the nostalgia the way he treats cowboy movies. New York smelled like an open sewer when horses handled transportation, because they left waste everywhere. Red Dead Redemption
-
Why Does Beautiful Code Still Matter for AI Agents?
David Heinemeier Hansson still sweats the details when agents change the Ruby that runs the 37signals business. For 25 years the economic reason was simple. A coherent, malleable architecture let a small team change the software quickly, without an exorbitant cost and without a pile of bugs every time one edit touched another. Beautiful code is easier to understand, simpler, and more malleable. That case assumed humans would do the edits.
The payoff of that sweat is diminishing, and how far it still matters is an open question. For the moment it does, because tokens are scarce. Anyone without an endless budget is token limited. A system an agent can change without relearning the whole conte
-
Why Do Detailed Prompts Hurt Coding Agents?
David Heinemeier Hansson says the system prompt shipped for Opus 5, and presumably for Fable as well, shrank by 80 percent. Boris, who works on Claude Code, shared that in an interview. The agent needed far less human instruction, and overly prescriptive humans were damaging the output. A hot stretch around agents.md had people micro-optimizing huge files of orders, one instruction after another.
He compares it to a boss who does not know the work and still dictates the method. A programmer sulks and writes worse code when the orders fight better judgment. An agent is no different. The skill he is practicing is to stop over-specifying and get out of the way, leaving a high-level vision. The
-
Why Can Programmers Be Worse at Directing AI Agents?
David Heinemeier Hansson says that for a while his own programming experience was a deficit. He told agents to implement the path he would have taken, and they were very good at obeying. In the first agentic stretch, about three months, that felt productive. He used everything he knew to prescribe the work. He was late to the next stretch, where describing the outcome and the problem produced a better solution than a programmer's prescribed path.
He will say it outright: there are problems for which programmers are worse than non-programmers. A lot of programmers are not very good product managers, and software is product management. What should it do, for whom, how should it look, what shi
-
Why Can Rust Be Ugly to Humans and Useful to Agents?
David Heinemeier Hansson hates reading Rust and still ships applications written in it. He calls it the ugliest programming language invented in probably the last 40 years, the opposite of Rails, and painful enough that looking at the code feels like acid in his eyes. He has written Rust anyway, and C++ as well, on vibe-coded projects of his own.
The output is why he puts up with it. Memory safety, a real systems language, and high efficiency are incredible, so Rust can be repugnant for a human to read and a wonderful platform for agentic engineering at the same time. He expects those two judgments to coexist. The agent can produce the code. He does not have to enjoy staring at it.
-
How Fast Did the Omarchy Plugin Marketplace Grow?
David Heinemeier Hansson says the Omarchy plugin marketplace reached 330 plugins in three days. He has never seen growth like that on any project he has been part of, or participation that broad, or so many people creating software that matters to them and is still usable by others, that quickly.
The plugin system lets someone extend the operating system and rewrite the interface, its panels, and its features by cloning tools that ship in the box. Omarchy ships with a calendar. A click on the clock pops it up, and that calendar does not consume appointments or speak iCal. About 17 implementations of a calendar that does already exist. The speed comes from skills Omarchy ships with, which te
-
Why Did Talk-to-code Fail for 50 Years?
David Heinemeier Hansson says the last 40 or 50 years of programming kept promising that ordinary people could write software just by talking. Fourth-generation languages were supposed to do it. Lisp was presented that way. Smalltalk was presented as an environment where regular people would build their own applications. None of it arrived in the sense that matters now. What people actually got was Microsoft Access databases and Excel spreadsheets, useful end-user tools and nothing like the output he is shipping.
He has been programming for 25 years, and the change he trusts is the one he can see: the output, the acceleration, the quality, and then the fact that it ships. Someone who has no
-
How Fast Can AI Agents Add an Omarchy Feature?
David Heinemeier Hansson started Omarchy last summer, between sessions at the 24 Hours of Le Mans, after too many videos about Linux got him bitten by the bug. An earlier attempt sat on Ubuntu and was fine. Omarchy started seven layers deeper in the stack, so the ambition was different, and he still wrote the early bash scripts by hand. A couple of versions had partial agent acceleration. Three months ago it went full throttle, at 100 percent. Agents write everything. He steers.
He can look at a feature in Windows, on the Mac, or on another Linux system, say he wants it, and the agents deliver. Most features arrive in five minutes, a few in 20, and a wild one takes two hours. The rush is st
-
How Does Omarchy Review a Flood of Pull Requests?
David Heinemeier Hansson says Omarchy now takes contributions from people who could not have sent them before. He has run the project for a little over a year. In the last three months, working on Quattro, he merged over a thousand pull requests. Quite a lot came from people who are not classical programmers, or from programmers outside Linux distribution work. Agents let them submit the idea, and he cherry-picks the best. About 400 pull requests are still unmerged, roughly double the count from a week earlier.
He has not been reading every request himself for some time. Agents review the queue and summarize whether something is ready for a human merge decision. Wrong, duplicated, and bad r
-
Can One Person Build Premiere for Linux?
David Heinemeier Hansson says one person can now build Premiere and Photoshop for Linux. He still reaches for Windows and the Mac when he needs video editing, and he treats that gap as something to fix rather than a reason to wait.
Tens of thousands of people who use Adobe Premiere or DaVinci already share the same frustrations. The complaints are public, on forums and on Reddit. What slows the incumbents is the same bureaucracy he describes inside large companies: management, meetings, and a product that has gotten so good that changing it is hard. He calls that the classic innovator's dilemma.
His side of the argument is impatience. Impatience is the first step of doing it yourself, and
-
Did Agents Write the Quattro Release of Omarchy?
David Heinemeier Hansson has not written, by hand, any of the code that shipped in Quattro. That is the Omarchy version he had been building for about three months, and it had just dropped. Agent acceleration on that version was near 100 percent almost from the start, and for the last two months it was 100 percent. He reviewed the shape of all of it and the individual lines of anything critical in the model layer. He did not review a lot of the UI or the auxiliary code, and he did not write the new functionality entirely by hand.
Basecamp and Hey were harder. Those professional products have large code bases and lots of users. Basecamp 5 was the first 37signals product that was really agent