It’s still early days in the AI revolution, but AI is already playing a big part in building more AI.
Boris Cherny, a Lead Engineer at Anthropic, shed light on the surprising extent to which Anthropic leverages its own AI to develop and refine the very codebase of Claude itself. He said that as much as 80 percent of the code for Claude Code, Anthropic’s CLI agent, is written by Claude Code itself.
“Probably near 80. Yeah, it’s very high,” Cherny said on the Latent Space podcast when asked about how much of Claude Code’s was written by Claude Code. But he did add that they had a human code review process too. “We do a human code review,” he added.
Cherny added that there were some parts that Claude was good at, and others which needed human intervention. “I think some of the stuff has to be handwritten, and some of the code can be written by Claude. And there’s sort of a wisdom in knowing which one to pick and what percent for each kind of task.”
Describing their typical workflow, Cherny explained how Claude often takes the first pass at coding tasks. “So usually where we start is Claude writes the code, and then if it’s not good, then maybe a human will dive in. There’s also some stuff where I actually prefer to do it by hand. If it’s intricate data model refactoring or something, I won’t leave it to Claude because I have really strong opinions, and it’s easier to just do it and experiment than it is to explain it to Claude. So yeah, I think that that’s how to maybe 80-90% Claude-written code overall.”
Cherny’s comments are an indicator of how AI is not just a product but an increasingly integral part of the development process itself. This “AI building AI” paradigm, even in its current AI-assisted form, has profound implications. It suggests a potential exponential acceleration in the pace of AI advancement, as models become more capable of contributing to their own evolution and optimization. In a fiercely competitive AI landscape, such efficiency gains derived from AI co-piloting its own development could be a significant competitive advantage.
This practice also highlights a significant shift in the role of human software engineers. While Cherny emphasizes the continued necessity of human oversight, critical review, and direct intervention for complex or “opinionated” tasks, the bulk of initial code generation can be offloaded to AI. This transforms the engineer’s role more towards that of an architect, a meticulous reviewer, and an expert prompter, guiding the AI and refining its outputs, rather than writing every line of code from scratch. The “wisdom in knowing which one to pick,” as Cherny puts it, becomes an even more crucial skill in this new era. As AI models like Claude become more sophisticated, their involvement in their own creation is likely to deepen, further blurring the lines between tool and creator and heralding a new chapter in software and AI development.