According to an OSChina report, one remark by Katelyn Lesse, head of Anthropic's platform engineering team, at the ObservabilityCON conference hosted by Grafana Labs drew industry attention: "As far as our team and our system development and maintenance work goes, all of our code is written by Claude. It really is all of it."
The team is no peripheral department -- it owns Anthropic's core services, developer tooling and product infrastructure, and had about 200 staff as of this summer. In Lesse's description, the work of writing and maintaining system code has now been handed over to Claude across the board, and the engineer's role has been restructured accordingly: from typing code line by line to system design, verifying results and gating code quality. While migrating a large codebase from Python to Rust, the engineers' main investment was designing verification mechanisms to make sure behaviour was consistent before and after the move, rather than rewriting every line themselves.
What makes this case unusual is its completeness. Most companies still position AI coding tools as assistive efficiency boosts -- generating drafts, completing snippets, with engineers reviewing section by section; Anthropic has taken it to an organisation-wide, full-volume production mode and backstops quality with a verification system instead of manual line-by-line review. Technical reports the company published earlier also indicate that its use of AI coding tools is not confined to the platform engineering team.
For enterprises still drafting their AI development standards, this news offers an advance sample: once the code-generation step is fully taken over by models, the competitive focus moves to requirement definition, system design, verification and quality gates -- the parts people are better at. Laying out test infrastructure, review process and the auditability of AI output ahead of time is worth more investment than arguing over whether AI should write code at all.
Source: OSChina