Most of the code. Written by Claude.

On August 17, Anthropic released its second Responsible Scaling Policy risk report. Buried in the report, they mentioned almost in passing: Claude now writes "most merged code" in Anthropic's own production codebase.

No launch event, no blog post blitz, not even a social media fanfare—it sat quietly in a paragraph of a risk report. But the weight of that sentence may be heavier than any AI coding product launch this year.

The Bootstrapping Tipping Point Is Here

First, let's clarify what "most merged code" means.

This isn't "assisting with coding" or "completing snippets" or "generating drafts that humans edit"—it's that most code merged into the production repository is written by Claude. This code runs in Anthropic's own products, serves millions of users, and handles real traffic and requests.

What does that mean? It means AI coding tools have crossed a critical threshold: from "assistant tool" to "primary contributor".

For two years we've debated how much code AI can write and whether it can replace programmers, but the framework was always "human-led, AI-assisted"—the programmer was the driver, AI was the co-pilot hitting the gas. Now Anthropic is telling you: on their own team, AI is already the primary driver, and humans are more like the co-pilot doing supervision and review.

Even more interesting is another line in the report: Anthropic noted that AI has significantly accelerated internal research and engineering, while its old task-based evaluation systems are beginning to show signs of saturation.

"Evaluation systems saturating"—translated into plain English: the tests and benchmarks we use to measure AI capability can't keep up with the speed of AI evolution anymore. When AI is writing code to improve itself, its progress is continuous, real-time, and non-linear—not "ship a new version, get an upgrade" like before.

The Role of Software Engineers Is Being Rewritten

Anthropic isn't the first company to have AI write its own code.

Over the past few months, we've seen signal after signal: the Cursor team says most new features are built using Cursor; the Midjourney team says their rendering engine is mostly AI-generated; many startups already run on a "1 engineer + 10 AI agents" configuration. But Anthropic is the first top-tier foundation model company to formally acknowledge that "most code in our production codebase is written by AI."

The psychological impact of this on the entire industry is enormous. If even Anthropic—one of the world's top AI companies with top-tier AI research and engineering teams—is using AI to write production code at scale, what excuse does any other company have not to follow?

That doesn't mean programmers are going extinct. But the role of programmers is fundamentally shifting: from "people who write code" to "people who review code, design architecture, define requirements, and enforce quality." Just as compilers meant programmers no longer wrote assembly by hand, just as high-level languages meant programmers no longer managed memory manually, the rise of AI coding means programmers no longer need to implement functionality line by line—but systems thinking, architectural ability, and product judgment become more important than ever.

The-AI-writing-AI loop has started. Code written by Claude makes Claude stronger; a stronger Claude writes better code. Where does this positive feedback loop end? Nobody knows. But one thing is certain: the software industry will never go back.

See you tomorrow.