Engineers say Claude Code now writes 90% of Anthropic’s production code, shifting staff to oversight

Anthropic’s own materials now support the broad direction of a claim that spread across social media this weekend: the company says Claude Code writes most of its internal code, and senior executives have described the share as above 90%. The most specific public attribution appears to come from Anthropic Chief Financial Officer Krishna Rao, whose remarks on a recent podcast were recirculated in press reports published between May 14 and May 17. Anthropic’s product page for Claude Code does not repeat the 90% figure, but it says “the majority of code is now written by Claude Code” and that engineers focus on architecture, product thinking and orchestration. The episode matters because it offers a rare public description of how an AI company says it is using its own coding agent inside production workflows. Anthropic has also published separate research and product material showing longer autonomous Claude Code sessions, broader use of auto-approval, and internal workflows built around review, testing and intervention rather than line-by-line manual drafting. (anthropic.com) ### Where did the “90%” number come from? Krishna Rao is the named Anthropic executive tied to the 90% figure in multiple reports published over the past several days. The Economic Times and Business Today both reported that Rao said on Patrick O’Shaughnessy’s “Invest Like the Best” podcast that “90 plus percent of our code is actually written by Claude Code.” (anthropic.com) Anthropic’s own website confirms Rao’s role at the company, though not that specific quote. A May 21, 2024 company announcement said Rao joined Anthropic as chief financial officer after finance and operating roles at Fanatics, Cedar and Airbnb. ### What has Anthropic itself confirmed on the record? Anthropic’s Claude Code product page makes the company’s clearest direct statement. (economictimes.indiatimes.com) The page says “the majority of code is now written by Claude Code” and adds that engineers now focus on “architecture, product thinking, and continuous orchestration,” including managing multiple agents in parallel and making decisions about what gets built. (anthropic.com) Anthropic has also described internal use cases in a July 24, 2025 blog post on Claude’s website. That post said product engineers use Claude Code as a “first stop” for programming tasks, while other teams use it for debugging, codebase navigation, automated pull request comments and writing tests. ### Does this mean engineers stopped coding altogether? Anthropic’s public materials describe a change in workflow, not the disappearance of human review. (anthropic.com) The Claude Code page says the system reads codebases, makes changes across files, runs tests and delivers committed code, while engineers provide direction and decisions. Anthropic’s research on agent autonomy points the same way. (claude.com) A February 18, 2026 paper said experienced Claude Code users increasingly let the tool run autonomously, but still interrupt it when needed, and that on complex tasks the agent pauses for clarification more than twice as often as humans interrupt it. ### How much autonomy is Anthropic saying these tools have? (anthropic.com) Anthropic’s February research paper said software engineering accounted for nearly 50% of agentic activity observed across Claude Code and the company’s public API. The same paper said the length of time Claude Code worked before stopping had nearly doubled in three months, from under 25 minutes to over 45 minutes, among the longest-running sessions. (anthropic.com) Anthropic has paired that push with messaging about controls. A March 2026 post on Claude Code auto mode said users approve 93% of permission prompts, and the company said it built classifiers to automate some decisions while trying to reduce approval fatigue. ### What can readers verify right now? Anthropic’s product page, available now, says the majority of its code is written by Claude Code and outlines the oversight-heavy engineering model the company says it uses internally. (anthropic.com) Anthropic’s research page, also live, gives current data on Claude Code autonomy and user intervention patterns. Anthropic is also still expanding the product’s commercial footprint. (anthropic.com) A company announcement from 2025 said Claude Code was being added to Team and Enterprise plans, and Anthropic’s current site continues to position Claude Code as a product for external developers and businesses as well as internal teams. (anthropic.com 1) (anthropic.com 2)