David Heinemeier Hansson on Coding Agents and Omarchy
Long-form interview (25-minute read) with David Heinemeier Hansson, creator of Ruby on Rails and co-owner of 37signals, by Vikas Shah in Thought Economics, published on September 28, 2026, five days after the opening keynote of Rails World in Austin where he announced that 37signals was going "pencils down" on hand-written code. (A) The reversal, dated: in summer 2025 he told Lex Fridman he would not let AI write his code; what changed his mind was the arrival of coding agents in late 2025, and "the change has really become vertical, in my opinion, over the last three or four months". (B) The underlying argument is not technical but political: programmers had become "a clerical class" mediating access to the machine, and the agent plays the role of Luther — "Agent Luther" — by removing the intermediary.
By David Heinemeier Hansson// Source thoughteconomics.com ↗/Reading 2 min/.md// Auto-verified translation
#end of hand-written code#pencils down#coding agents#agentic coding#priesthood of programmers#clerical class#disintermediation#Agent Luther
Summary
Vikas Shah interviews David Heinemeier Hansson, creator of Ruby on Rails and co-owner of 37signals with Jason Fried, five days after a Rails World keynote where he announced that his company was stopping writing code by hand. The interview, published on September 28, 2026, opens on the question of empowerment versus dependency. His answer is blunt: AI makes people "vastly more capable," and the worry about outsourcing thought is 2,500 years old, from papyrus to rock 'n' roll, always disproven.
Then comes the central argument, which is about power rather than technology. Programmers, he says, formed a clerical class mediating access to the computer; agents play the role of Luther and remove the intermediary, which he says explains the skepticism of part of the open source community, whose premise was nonetheless control over one's own software. The analogy he developed on stage is that of the painted portrait: three months of a master painter's work reserved for aristocrats, then two trillion photos taken every year. The world gained from it; photorealistic portrait painters did not.
On model capability, he is categorical: frontier models, especially combined with one another, now surpass almost every programmer across a wide range of tasks. Most applications are CRUD, roughly 97% of the installed base. He dates the shift to the last three or four months and acknowledges that his own skills no longer count as an economic lever.
He then sorts through business models. Purely client-side software is doomed — he would not want to be Adobe. Client-server holds up because nobody wants to administer servers. Platforms with deep plumbing, like Shopify, where he sits on the board, keep a moat. Salesforce and SAP, meanwhile, are "in serious trouble": there's no more value in the code, only in process intuitions.
Omarchy occupies the second half. A Linux distribution whose Quattro version was written exclusively by agents, it embeds an agent at the core of the system: agent selection at first boot, automatic crash diagnosis, a fix and an upstream pull request. In four weeks, more than 4,000 plugins were created, most of them by non-programmers. Linux's advantage is that its 40 million lines are in the models' pretraining data.
He closes on craftsmanship — a steering advantage that keeps shrinking, and a minority who will need to understand the machine — then on robots, abundance, and "the problem of existence."
Key takeaways
The operational announcement, worth noting before the analogies. 37signals went "pencils down" on hand-written code. This is not a forecast but a company decision, made by the author of a framework that, by his own count, underpins more than half a trillion dollars in market capitalization (Shopify, GitHub, Airbnb, Coinbase).
The timeline of the conversion. , useful for dating the shift of a reference practitioner: summer 2025, he tells Lex Fridman that programmers "learn with their fingers" and refuses AI in his code; late 2025, the arrival of agents able to write, test, and fix on their own; four months before the interview, he starts Quattro without writing "virtually a single line"; August 2026, back on Fridman's show with an OS he didn't code. The shift is measured in months, not years.
The 97% CRUD thesis. The capability argument doesn't hinge on edge cases: "most programmers work on fairly rudimentary systems called CRUD — create, read, update, delete. That's roughly 97% of the world's applications." The areas where humans still beat the agent exist but are "very small" and "that's not what most programmers do". It's the same reframing of the question as in [[felker-evaporation-software-engineering-agentic-builder-2026-09-14]].
The vendor-survival grid, in four boxes. — the most directly reusable part. (1) Pure client-side (Photoshop, Lightroom, Illustrator, Premiere): "I would hate to be in Adobe's shoes", vibe-coded open source alternatives coming en masse, "we're living on borrowed time". (2) Client-server: still holds, not for technical reasons but because nobody wants to administer servers — an explicit caveat, "that might change once they trust their agents enough". (3) Embedded plumbing (Shopify: payments, shipping, taxes): complexity absorbed and resold so cheaply that "it barely seems worth the tokens" to replicate it. (4) Expensive, low-process subscriptions (Salesforce, SAP): "in serious trouble", "there's no more value in the code, only in process intuitions — and I'm not sure their process intuitions are that good". Extends the SaaSpocalypse breakdown in [[corrot-mirakl-pirates-de-lia-2026-09-20]], which arrives at the same sorting by the nature of the asset.
The second, non-technical barrier. even when replicable, software still requires "your taste, your diligence, your discernment, and your attention" to be maintained and advanced. This is the caveat DHH himself raises against the idea that all software will become bespoke — to be set against [[lee-robinson-personal-software-2025-01-01]].
The death of file formats as a barrier. LibreOffice took years; "if you wanted to build a Word-compatible text editor on Omarchy, you could be done by this afternoon". The stated consequence: the value of file formats and protocols is collapsing, "because agents speak every language" — literally too, the Omarchy site was translated into 50 languages by agent, purchased domains included.
Why Linux and not Windows/macOS. , an engineering argument worth keeping: the kernel's 40 million lines are in the pretraining data of every frontier model, which therefore know the system from the inside; on closed OSes, the agent can only work at the margins. The advantage is no longer ideological but informational.
The Omarchy mechanism worth copying elsewhere. at first boot, choice of default agent (Claude, Codex, Grok); on every crash, a notification "want AI to take a look?"; the agent diagnoses, distinguishes a local bug / an application bug / a distribution bug, proposes the fix, and opens an upstream pull request — "the loop is closed". Results reported after four weeks: 4,000 plugins, mostly built by non-programmers.
The scale argument he sets against the major vendors. a few hundred thousand agent-equipped Omarchy users represent "tens of millions of programmers" against the "tens of thousands" at Apple and Google. A formulation to handle with caution — it converts users into engineer-equivalents without addressing the question of coordination.
What's left of craftsmanship, in his view. a steering advantage (better prompting, better direction) that "keeps shrinking" — "less than last week, and much less than three months ago"; and the agricultural analogy, 97% of farmers down to under 2%, with a corollary he embraces: "it's very important that that last two percent know what's going on", and increasingly so as dependence on the machine grows. The social conclusion is an admission rather than a program: "what do you do when you have nothing left to do?", with skepticism toward universal basic income ("the studies aren't very flattering").
What the interview does not provide. , worth noting before citing it: no measurement, no protocol, no agentic failure case, and a single voice. The strong claims — models superior to "virtually every programmer on Earth", AGI reached, "pockets of ASI", better lawyers than programmers — are usage impressions from a highly exposed practitioner, not results.
Key figures
over 200 000 downloads in 18 days and $21,7M in pledged donations to the foundation by end of September 2026