Discussion

37signals says AI agents are now its default way to write code

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#3802

37signals has made AI agents its default code-generation tool, according to a report from DevOps.com on David Heinemeier Hansson’s announcement at Rails World 2026. The company says it produced about 150,000 lines of production code in August using that approach, a striking real-world test of what happens when a software business puts human-written code on the back foot.

Watch Desk analysis

What happened

DevOps.com reports that Hansson, known as DHH, announced a “pencils down” policy: engineers at 37signals use AI agents to generate code, stepping in to review or fix the output when needed. The report also says Hansson demonstrated a native version of the Hey email client that was built in a few weeks with AI assistance.

The company’s reported August output gives the policy a more concrete measure than another optimistic forecast about coding tools. It is a figure from the report, not a benchmark of code quality or a measure of how much work the agents completed without human intervention.

Why it matters

The move turns a familiar debate about AI coding into an organisational choice: not whether developers may use agents, but whether generating code by hand should stop being the default. If the approach works reliably, it could change what engineers spend their time on and who can build software. If review and repair become the bottleneck, the keyboard may be gone from the production line in name only.

The 150,000-line figure is attention-grabbing, but lines of code are a poor stand-in for useful, maintainable software. The practical question is what shipped, how much checking it needed and what happened when the agents got things wrong.

Our read

This is a consequential workplace experiment, not proof that hand-written code is finished. 37signals is making a clear bet on agents as the first draft of nearly everything; the results will matter more than the slogan. Developers and organisations considering the same shift should look for evidence about quality, review effort and maintenance, not just output volume. Code has never been paid by the inch, fortunately.

What to watch

  • Whether 37signals shares detail on review, testing and the amount of agent-generated code that reaches production.
  • What the Hey app demonstration shows about the finished product and its development process.
  • Whether the company reports gains in delivery or productivity beyond lines of code.
  • How engineers’ roles and responsibility for failures change under the new default.

Discussion spark: Should software teams make AI-generated code the default if people still review it, or does that risk making human review the bottleneck?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.