DHH: Approval Layers, Not Coding Speed, Cap AI Productivity

My notes

Summary

DHH argues large, established software companies aren’t seeing 10x productivity gains from AI agents because their bottleneck was never implementation speed. It is human bandwidth, approval layers, and lack of vision or taste. The real 10x to 100x gains only show up when a person interacts with agents directly, with no human intermediary slowing the loop down. He backs this with his own case: in roughly 3 months of working this way he became a “polyglot programmer,” shipping C++ apps despite being a lifelong Ruby and Bash developer.

Key Insight

• Implementation was rarely the bottleneck. Product managers, designers, VPs, and CTOs all wanting to shape a decision is where productivity dies, not the coding itself. Giving a big org agents doesn’t remove that layer, so throughput barely moves.
• Most orgs are bottlenecked on ideas, vision, and taste, not capacity. Microsoft has had “endless resources, endless programming capacity for decades” and that alone never produced compelling software, which is proof that raw implementation capacity was never the constraint.
• The 10x to 100x multiplier requires removing the human-to-human handoff entirely. You have to talk to the agent yourself. Three layers of approval plus agents is still slow, because the approval process is the actual bottleneck, not typing speed.
• Six months of this capability existing is not enough time to judge it. DHH frames “why hasn’t this changed everything yet” as itself a category error given how new agentic capability is.
• Concrete personal example: he became a polyglot programmer despite being Ruby-first with only some Bash. In the last 2 months he wrote C++ and shipped three applications in his own Linux setup.
• The “5% of a bloated app” pattern: he replaced Typora, a markdown editor he was using only because iA Writer isn’t available on Linux, by telling an agent to build a C++ and Qt clone of the roughly 5% of features he actually used. First working version took about 20 minutes, and he had fully switched over for all his essays within 2 days.
• Reframes the “rewrite everything in Linux” idea not as porting full apps like Photoshop or Premiere, but as each person rewriting their own narrow slice of functionality. As he puts it, we all use a different 5%, so what if we all just build our own 5%.
• Computing platforms are newly “in play” for the first time in roughly 40 years, since the desktop era began, because agentic coding lowers the cost of building single-purpose replacements for software that locks you into a platform.