DHH Goes Agent-First on Everything at 37 Signals

2026-04-08 The Pragmatic Engineer

agents future-of-work productivity interview business

How DHH Went from AI Skeptic to Agent-First Developer

David Heinemeier Hansson — creator of Ruby on Rails, co-founder of 37 Signals, and one of the most opinionated voices in software — has done a complete 180 on AI coding tools. In this in-depth conversation with Gergely Orosz on The Pragmatic Engineer podcast, DHH explains exactly what changed and why.

The autocomplete era was infuriating: “I found it as we’re trying to have a conversation. You won’t let me finish a sentence. You’re constantly trying. Was this what you meant? Was this what you meant? You’re like, shut the hell up.” DHH’s frustration with tab-completion AI (Copilot, Cursor) was visceral — experienced developers don’t want a tool guessing their next character. They want to think.

Then agent harnesses changed everything: The combination of Claude Code / Open Code as terminal-based agent harnesses plus Opus 4.5 (released November 2025) was the inflection point. “It produced code I wanted to merge without very much if any alteration and if I did want to do alteration I could tell it and it would remember and it would not make the same mistake next time.” The shift from autocomplete to autonomous agent execution — where the AI has tools, bash access, and internet — transformed the entire experience.

Agent-first means starting with the agent: DHH’s workflow has completely inverted. Before: open editor, write code, ask AI for help when stuck. Now: tell the agent what to build, review the output, make refinements. He runs two models in parallel — Gemini K25 in Open Code on top, Opus in Claude Code on bottom — reviewing diffs in Neovim as they come in.

The “mech suit” realization: “Running a bunch of agents feels less like being a project manager for agents and more like stepping into this super mech suit where suddenly I don’t just have two arms. I have 12.” This counters his own prediction from Lex Friedman’s podcast where he said he didn’t want to be a “project manager for agents.” The reality turned out to feel like a superpower, not delegation.

100 PRs reviewed in 90 minutes: Before the Omachi 3.4 release, DHH faced 250 pending pull requests. He pointed Claude at each PR URL and processed 100 in 90 minutes — some merged as-is, some rewritten by Claude in the correct project style, some closed. Work that would have taken days of manual review. “This would have been a week’s worth of work, days at the very least.”

5 Key Insights from DHH on AI and the Future of Software

The pie is exploding, not growing — 37 Signals is tackling projects they never would have contemplated. One senior developer optimized P1 (the fastest 1% of requests) from 4ms to 0.5ms across 12 PRs in days — a “vanity” project no one would have approved before agent acceleration made it essentially free.
Senior developers benefit most — The biggest AI acceleration accrues to the most experienced people who can validate agent output against production requirements. Amazon’s internal analysis pinned major outages on junior developers shipping unreviewed agent-generated code. The skill gap is widening, not narrowing.
“Peak programmer” may be here — DHH argues we may be nearing the end of the era where developers could command premium compensation simply by being the implementation bottleneck. More software than ever will be produced, but the constraint is shifting to taste, judgment, and knowing what to build.
Design and taste are becoming more valuable — At 37 Signals, designers are product managers who also write CSS and HTML. Agent acceleration is making this model more viable industry-wide, as designers can now implement their own visions end-to-end. Aesthetics isn’t a luxury — it’s a signal of correctness.
Ruby on Rails is having an AI renaissance — Rails is one of the most token-efficient ways to build web apps, making it ideally suited for agent workflows where context window and cost matter. Beautiful, readable code isn’t just for humans anymore — it helps agents produce better output too.

What This Means for Organizations Going Agent-First

DHH’s experience at 37 Signals — a 60-person company running for 22 years — offers a practical blueprint: the same team doing dramatically more ambitious work, not a smaller team doing the same work cheaper. The key insight is that agent acceleration doesn’t just speed up existing tasks; it unlocks an entire category of work that was previously too expensive to even consider. For organizations weighing whether to invest in agent tooling, the question isn’t efficiency — it’s ambition.