Build1 distinct publisher3 min readPublished
One programmer ran the same bet on two codebases and got opposite results, which is closer to a controlled test than anything else in the agentic coding argument, and the published dates make it harder to read than it needs to be.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
An agent working a feature ticket sees the repository, the tests and the prompt. It does not see the decisions that were made once and never written down: which layer may touch the database, which object owns a lifecycle, which duplication was deliberate. Local correctness is checkable, which is why each pull request passes even though nothing checks the union of pull requests. The publisher's framing is right on that point, that an agent can produce functioning code without understanding the years of accumulated decisions that make a large application maintainable [19], and Hansson says this is what showed up at 37signals as pull requests that read reasonably one at a time and needed human programmers to clean up [7].
Reviewing overall structure is cheap when the structure fits in one head. Hansson says he directed Quattro, inspected its structure and reviewed critical sections while hand-writing none of the shipped code [3], and the official release materials describe a unified Quickshell-based shell, dual-boot support and faster installation [13], so the agent work covered a shell, not a theme refresh. Copying that division of labour means reproducing the conditions attached to it: one owner holding the whole design, and a repository young enough that this month's choices are not constrained by old ones. Omarchy started in the summer of 2025 [12]. Basecamp 5 was not built that way, since designers were implementing features directly alongside programmers [6], and the codebase belongs to the company where Hansson built Rails before releasing it in 2004 [14].
"All the code shipped" is a provenance claim, and provenance is measured at the commit. It is compatible with a great deal of human work upstream of the commit: specifying the problem, choosing among proposed solutions, judging whether the result is coherent [5]. Hansson closes off the loose reading himself, defining vibe coding as asking an agent to build software without examining the implementation, and saying his own workflow still depends on technical judgment [10].
The dates need work. The account puts his change of mind at November 24, 2025 [9], the Fridman conversation's posting at August 26 [2], and the Quattro release at around August 21 [4]. Two months of agent-led implementation ending at that release places the work somewhere between late June and August [17], which puts the stated conversion after the release it is offered to explain [16]. One of those dates is wrong; a five-hour interview will not tell you which.
The softest figure is the one the publisher already flags: tens of thousands of downloads shortly after launch, on Hansson's own count, with no independently measured active-user number [11]. Installs say nothing about whether agent-written code survives its second year of changes. As evidence this is one participant's report on two codebases he owns, thinner than a study and more useful than another benchmark table, because it names the variable [18].
Ranked by verification strength, evidence, and original report placement.
David Heinemeier Hansson, creator of the Omarchy Arch Linux-based desktop distribution, told Lex Fridman that AI agents wrote all the code shipped in Omarchy 4.
The conversation with Fridman ran five hours and was posted August 26.
Hansson said he still directed the product, inspected its overall structure and reviewed critical sections, though he wrote none of Quattro's shipped code by hand.
Omarchy Quattro was released around August 21, according to the interview.
Hansson told Fridman that agents handled effectively all implementation during Quattro's final two months, and that his role moved toward specifying problems, choosing among proposed solutions and judging whether the result felt coherent.
During the final development sprint for Basecamp 5, 37signals allowed designers to use agents to implement features directly.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 31, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Omarchy Quattro pre-wires every coding-agent CLI and refuses to name a default1 distinct publisher
build
The 680 MB database that was really a 17 GB disk: self-hosted support platforms fail at month six1 distinct publisher
invest
A CEO's $1,000 weekend, and the auto-renew setting that made it possible1 distinct publisher
build
Four indexes, none of them covering: the 78-second page and the one index that fixed it1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One podcast, one outlet, no independent check
Agents wrote all of Quattro, agents dented Basecamp's architecture, tens of thousands downloaded it — all three come from the same man in the same interview, relayed by the only outlet covering it. The independent paper on the table confirms the wrong things: Omarchy's release notes describe a Quickshell shell and dual boot, 37signals' REWORK page confirms Hansson's title, and neither says anything about who wrote the code. Then the dates fight each other, with the conversion offered as cause landing after the release it explains.
A real distro, a real sprint, one hand on both
Something genuinely shipped: Quattro exists with a rewritten shell and dual boot, and its author counts tens of thousands of downloads — his count, no telemetry. The only account of this practice inside a commercial team is 37signals' Basecamp 5 sprint, and that one ends in remediation. So adoption is two instances, both belonging to the same person, one of which is the cautionary tale.
The headline outruns its own footnotes
'Agents wrote all the code' travels a great deal further than the conditions attached to it — an experienced product builder, a distribution he started months earlier, and a review pass he performed himself. The same interview reports programmers repairing what agents did to Basecamp 5. runtimewire.com does the honest thing and keeps narrowing the claim, but the narrowing is in paragraph fourteen and the reversal is in the first line.
The author of the distro is the witness for it
Hansson is selling nothing at a price, but he wrote the distribution, counts its downloads, and holds the thesis the story validates: that one experienced builder can direct far more software. A five-hour podcast is a friendly venue for that. Cutting the other way, the sharpest evidence against his own rhetoric — designers' pull requests bending Basecamp 5 out of shape — is a disclosure he volunteered about his own company.
Sure what was said, unsure what happened
We can be fairly confident of the transcript — the quotes, the workflow description, the Basecamp cleanup — and much less confident of the underlying facts. A single outlet, a single interview, an author-supplied download number, and a timeline that does not survive being read in order together mean the interesting comparison in this story is credible in shape and unaudited in detail.