DHH's new way of writing code: Agent-first
7 hours ago
- #AI in Development
- #Product Design
- #Software Engineering
- David Heinemeier Hansson (DHH) has shifted from manually typing code to using an agent-first approach with AI tools, significantly changing his development workflow in the past six months.
- DHH uses a setup with two AI models: one fast LLM (typically Gemini 2.5) and one slower, more powerful model (usually Opus), integrated with tmux and NeoVim for reviewing diffs via Lazygit.
- He believes that beautiful code and products signal correctness, linking aesthetics to functionality, and emphasizes that strong design opinions lead to better software.
- Ruby on Rails is experiencing a Renaissance due to AI, as it is token-efficient, suited for agent workflows, includes built-in testing, and produces human-readable code for easier review.
- Using AI agents allows tackling previously unconsidered projects, like optimizing the fastest 1% of requests at 37signals, reducing response times from 4 milliseconds to under half a millisecond.
- Senior engineers benefit more from AI than juniors, as they can validate agent output for production readiness, while juniors may require supervision, a conclusion also reached by companies like Amazon.
- 37signals has a high designer-to-engineer ratio (1:2), with designers acting as product managers and implementers who understand both design and implementation, similar to jewelry designers knowing materials.
- AI agents could make 37signals' designer model an industry standard, empowering designers to implement their vision directly and converge towards small, multidisciplinary teams.
- Command Line Interfaces (CLIs) are seen as the ultimate AI interface, validating the Unix philosophy by allowing agents to chain tools together, with 37signals building CLIs for all products.
- The two-month product development cycle described in 'Shape Up' is now outdated due to AI acceleration, requiring a rewrite of the methodology to reflect faster timelines.
- DHH maintains a non-negotiable eight hours of sleep to avoid burnout, despite the intoxicating dopamine loop from shipping with AI agents, emphasizing work-life balance.
- The episode discusses the uneven impact of AI, amplifying senior engineers while creating challenges for junior developers, and explores what this means for the future role of software engineers.