Generated by Codex with GPT-5

What happened

The Pragmatic Engineer surfaced this April 8, 2026 piece, and the original post is DHH’s new way of writing code.

Gergely Orosz interviews David Heinemeier Hansson about a notable reversal. Six months ago, DHH said he did not want AI writing his code. Now he describes an agent-first workflow built around terminals, tmux, Neovim, and multiple models running in parallel. His argument is not that the principles of software craft changed. It is that the tools improved enough to make delegation worthwhile. Autocomplete used to feel annoying and low-value; newer agents can now produce code that he is comfortable reviewing and sometimes merging with minimal changes.

The piece also widens from DHH’s personal workflow into a broader claim about the shape of software teams. He argues that AI rewards engineers who can judge quality, designers who can implement their own ideas, and organizations that already work in small, opinionated, high-trust groups.

The simple version

This is really a story about where software value is moving.

DHH is spending less time typing every line by hand and more time setting direction, reviewing diffs, choosing tools, and deciding what good looks like. In his framing, AI does not replace taste, correctness, or technical judgment. It shifts them closer to the center of the job.

Why it matters

• The piece makes a sharp distinction between generating code and evaluating it. DHH argues that senior engineers benefit more because they can tell when agent output is solid, risky, or subtly wrong.
• It offers a concrete reason some stacks may age well in the AI era. DHH sees Ruby on Rails as unusually compatible with agents because it is compact, readable, and test-friendly, which makes generated code easier to inspect and verify.
• It suggests that software teams may get smaller without becoming slower. If engineers can offload more implementation work and designers can build more directly, companies may need fewer handoffs and shorter delivery cycles.
• It gives the command line a strategic role. DHH argues that CLIs are ideal interfaces for agents because they expose clear actions that can be chained together across tools like GitHub, Sentry, and internal systems.
• It also adds an important constraint: speed is intoxicating, and burnout is a real risk. Even in an agent-heavy workflow, DHH treats sleep and human judgment as non-negotiable.

Takeaway

The most interesting part of this piece is not that DHH now likes AI tools. It is that someone so closely associated with handcrafted software thinks the center of gravity has shifted from writing code line by line to directing and auditing machine-generated work.

That does not make engineering less demanding. If anything, it raises the premium on taste, review discipline, strong interfaces, and clear tests. Teams that already have those habits may get dramatically faster. Teams that hope AI will substitute for them may just produce more code with less control.