37signals has stopped treating hand-written code as part of its routine development process, Ruby on Rails creator David Heinemeier Hansson said at Rails World. The software company now relies on AI coding tools, with engineers writing code by hand only when those tools fail to deliver the needed result—a shift that puts the balance between development speed and human oversight in focus.
At 37signals, manual coding is the fallback
“We don’t write code anymore,” Hansson said in his Rails World keynote, describing the approach as “putting down the pencil.”
At 37signals, engineers can step in and write code if an AI agent cannot produce what they need. But the next step, Hansson said, should be to “fix the machine,” rather than make manual coding a regular part of the workflow.
He framed the policy not as a sudden change in how the company works, but as recognition of a shift already under way. In his view, the central question for software development now is how to make better use of AI’s benefits; other concerns rank below that priority.
Developers remain divided on AI coding tools
Tools including OpenAI Codex, Anthropic Claude Code and Cursor have rapidly changed how some developers approach their work over the past year. Some describe AI coding agents as exciting enough to keep them working late, while others say AI-assisted coding can make the process feel isolating or exhausting.
In a survey of more than 150 software engineers conducted in January 2026, over 45% said they were “keeping up” with AI tools. Another 17.5% said they did not use AI code editors at all. The results suggest that AI tools have entered many developers’ workflows, but adoption remains uneven.
Hansson also asked the Rails World audience who still wrote a lot of code by hand each week. Judging by the show of hands, he estimated that “maybe five” people did. The exchange echoed 37signals’ approach, but does not by itself indicate how the wider software industry uses AI.
For software companies, AI coding tools may affect more than the amount of code engineers type themselves: they can also change how work is divided and how code is maintained. Hansson’s approach is to have people intervene when AI cannot handle a task, while focusing on improving the tools. Developers’ differing experiences suggest there is no settled view yet on whether that model works across teams.