David Heinemeier Hansson, the Ruby on Rails creator and technology leader at 37signals, once voiced a deeply personal reservation about AI-assisted programming: he said he avoided it because he could “literally feel competence draining out of my fingers.”¹ This spring, 37signals’ effort to use AI heavily on Basecamp 5 also produced “mixed” results.¹

Yet at Rails World, Hansson presented that experience not as a warning but as a reason to commit harder. He said 37signals had “gone, pencils down” on manually written code, making hand-coding an “exceptional state” reserved for cases where an AI agent cannot deliver.²

The practical rule is blunt: when an agent gets it wrong, engineers may “get the old pencil out and dot it down for them” — then “fix the machine.”² Hansson framed the policy as recognition of an industry change already underway, rather than a speculative bet. In his view, the only serious question is how to capture AI’s gains; other concerns sit lower in the hierarchy.²

That confidence is not universal. The same account of the shift describes developers staying up late to build with coding agents, treating tools such as Codex, Claude Code and Cursor as exhilarating collaborators. Others find AI coding lonely or exhausting.²

A January poll of more than 150 software engineers illustrates the fault line: more than 45% said they were keeping up with AI tools, while 17.5% said they were opting out of AI code-editing tools entirely.² Hansson’s answer is to stop debating whether agents belong in the workflow. For a sizable minority of developers, however, the pencils are not down yet.