DHH says 37signals stopped writing code by hand, and he wrote 150,000 lines in August

At Rails World 2026, David Heinemeier Hansson said agent-generated code is now the default at 37signals, and hand-writing code is the exception kept for fixing the workflow when an agent misfires. He put a number on his own month: about 150,000 lines of production code in August 2026, against a 21-year average near 30,000 a year. He also said HEY is being rebuilt with six native apps and a Rust mail backend, citing 99% less CPU for work still in progress. (The Pragmatic Engineer)

Line counts are a bad metric and DHH knows it, which is partly why he used one. 150K lines in a month is 5,000 a day, every day, including weekends. Nobody reads 5,000 lines a day with care. So the claim underneath the number is that reading every line stopped being the quality mechanism, and something else took over.

That's where I get uneasy, and it's not a hype objection. Today's other findings are all about that "something else." The SpecHarness result says models claim completion 28.7 to 37.9 points more often than the evaluator agrees. arXiv 2609.29744 published the full development history of a 21,000-line Python tool with no human-written code or tests, and the AI-written test suite later caught a real error in 14.3% of generation events, with about one in four to five interactive responses containing at least one factual error. arXiv 2609.28850 found agents reproduce 41% of NeurIPS 2025 papers even at the tier where code, data and weights are all released, and the single most common failure across 400 runs was writing the method without checking any intermediate number against the paper.

37signals has something most teams don't: a two-decade test suite, a product with known behavior, and a person at the top who can tell in ten seconds whether Basecamp feels wrong. That's the verification layer. It isn't code review, it's a body of tests plus extremely expensive taste. DHH can go agent-default because he already built the thing that catches agents.

If you don't have that, the honest version of this move is to build the verification layer first and the volume second. I mean that literally: the deterministic checks, the property tests, the golden files, the thing that goes red without a human reading a diff. Anthropic's own engineering post on making claude.ai 3.1x faster describes the same ordering, they built deterministic lab metrics first (Valgrind instruction counts under Node's --predictable, V8 call counts, React commit counts, DOM mutations) and verified each tracked wall-clock time before letting Claude optimize against them.

I'm not arguing DHH is wrong. He shipped the artifact. I'm arguing the transferable part of his result is the 21 years of tests, not the 150,000 lines.