David Heinemeier Hansson: DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux

DHH describes his change of practice as a response to different tools rather than a reversal of his values. Earlier autocomplete interrupted the craft he enjoyed, while chatbots were useful tutors. His personal dividing line was trying Opus 4.5 in late November 2025: the output began to resemble code he would write, and tool use, self-checking and computer operation turned intelligence into usable work. This is an account of his experience, not a controlled comparison of model intelligence.

He separates successive stages: directing one agent, coordinating parallel subagents, and describing a problem while the system proposes a route. Omarchy brought him to entirely AI-written implementation, but Lex immediately distinguishes ordinary web applications from operating systems and safety-critical software. DHH's confidence is explicitly grounded in the domain he currently works in. His security examples argue that AI can also inspect difficult code; they do not establish that every domain can dispense with human review.

Basecamp supplies a counterexample to uncomplicated acceleration. Letting designers generate changes produced pull requests that could look defensible individually yet collectively damaged architectural coherence. Humans had to clean up the result. DHH limits the lesson to substantial existing codebases whose architecture matters, and notes that this episode occurred in February, before subsequent improvements. He also rejects the assumption that human-written code is automatically orderly.

His broader bottleneck is organizational: approval layers, communication, unclear ideas and weak product judgment can absorb implementation gains. Personal software offers a different route. Rather than reproduce every feature of a large application, he recreated the small subset of Typora he used, producing Omawrite in C++ and Qt and switching within two days. Sharing such tools can widen their audience, but Lex's question about excellence and maintenance remains important: publication alone does not resolve stewardship.

DHH argues that agents can improve contribution hygiene: complete reports, explanations, tests and review are tasks they can perform when instructed. He reports merging more than a thousand Omarchy pull requests during Quattro's development, including ideas from people outside Linux development. His answer to review overload is both selective rejection and agent triage, with humans retaining the merge decision. That reframes the maintainer's job rather than making every incoming contribution valuable.

He no longer assumes that useful ideas must originate with humans, and describes his excitement as a reaction to working, shipped software. Omarchy is his concrete exhibit for claims that otherwise risk sounding detached from results. Lex raises the drug-like pull of abundant capability and divided attention; DHH answers that a singular mission, building his ideal computer, channels the energy. Their difference is about how the same capabilities affect attention, not whether excitement itself proves AGI.