Anthropic CEO Dario Amodei revealed this week that a growing portion of his engineering team no longer writes code manually — they direct Claude to write it, then review and edit the output.

"I have engineers within Anthropic who don't write any code," Amodei said in a widely-shared clip. "They just let Claude write the code and they look it over."

More significantly, he added that much of Claude's own development now runs through the same loop. "Writing code at Anthropic means designing the next version of Claude itself — so we essentially have Claude designing the next version of Claude."

50+ Features in 52 Days

The claim is backed by a concrete output metric: the Claude team shipped more than 50 major features between early February and the end of March 2026, a pace that Amodei attributes directly to the AI-driven development workflow. Productivity benchmarks internally show a 27% uplift for developers using the tool.

Context

This isn't a theoretical AI-safety discussion — it's a description of how a $19B-revenue-run-rate company operates today. Anthropic's annualized revenue grew from $1B to $19B over roughly 15 months, with coding tools cited as the primary driver.

The self-improvement loop Amodei describes — AI building better AI, humans supervising — is now operating at commercial scale, not in a research lab.

What It Means

Human engineers at Anthropic have shifted from authors to editors. The bottleneck is no longer coding speed but judgment: deciding what to build, reviewing what the model produces, and catching errors it doesn't catch itself.