Agentic Engineering Is Now Writing All the Code

DHH says agentic engineering just flipped from autocomplete to autonomy, and he is shipping at 100% AI-written code.

The interview is YouTube episode #501 of the Lex Fridman Podcast, recorded 13 months after their previous conversation.

What changed for DHH on AI?

He was a skeptic of autocomplete and chatbot modes a year ago, then Opus 4.5 on November 24, 2025 became his dividing line.

He tried it on November 26 and found output uncannily close to his own Ruby, not just solving tasks but taking the path he wanted.

The jump was not only model intelligence but harnessing: tool use, self-checking and instrumentation that turned demos into mergeable code.

By early spring sub-agents cut task time to a fifth or a tenth by splitting work across eight workers, though he still had to drive.

This summer with Opus 5, Fable and GPT Sol the role reversed: he states the fuzzy problem and the system picks the route.

Why does Omarchy prove the point?

Omarchy is his Arch-based Linux distribution built around the Hyprland Wayland tiling compositor, now at version Quatro launched last Friday.

He says the last two months of Quatro shipped with zero lines written by hand, with critical model-layer code reviewed and much UI code not even inspected.

That contrasts with Shopify (NYSE:SHOP) CEO Tobi Lütke’s early push and DHH’s own Basecamp 5 sprint from February, where designer-led vibe coding produced PRs that wrecked architecture and required manual cleanup.

He built Omawrite, a Typora and iA Writer replacement in C++ and Qt, to a first version in about 20 minutes and fully replaced Typora within two days.

He hates Rust as the ugliest language of the last 40 years and yet shipped Rust apps because agents absorb the syntax pain while delivering memory safety and speed.

What does this mean for software teams?

DHH argues large teams are bottlenecked on communication and ideas, not implementation, and Microsoft (NASDAQ:MSFT) shows that having tens of thousands of programmers never guaranteed great software.

He says direct interaction with agents delivers 10x to 100x gains, sometimes 1000x, but inserting human approval layers kills the bandwidth gain.

That fits the classic innovator’s dilemma he cites: incumbents tuned for the old process cannot pivot super tankers fast enough.

StartupHub.ai data shows You raised $80M in its 2023 Series A, and the firm tracks You alongside Perplexity AI, Google, Lucidworks and others competing to own the AI answer layer where agentic coding now plays.

For open source, he merged over 1,000 pull requests in three months on Omarchy, with about 400 still unmerged, and saw 330 plugins appear in three days after shipping agent skills for extensions.

He now lets agents triage PRs and validate fixes in a VM, reviewing only the pearls, and says he prefers agent PRs because they follow instructions, include tests and comments, and are painless to reject.

He also warns models are now strong enough to chain multiple small vulnerabilities into remote code execution, so security review still matters even as AI finds flaws better than most humans.

The GPS analogy captures his current trust: early GPS might drive you into a harbor, but now cars drive themselves on trusted routes in web CRUD and his Linux work.

If agents stay this capable, the scarce skill shifts from loops and variables to product judgment about what to build and for whom.