The most AI-native engineering teams on earth are at Anthropic, and a reporter got inside access. Four people spoke on record: Katelyn Lesse, Head of Engineering for Claude Platform; Jarred Sumner, creator of Bun, now working on Bun and Claude Code; Thariq Shihipar, spanning Claude Code engineering and education; and David Hershey, Applied AI, working directly with customers like Cursor, Cognition, and Perplexity. This is primary sourcing from the lab building the tools everyone else is using.
Three findings anchor the piece. First, a new agent infrastructure primitive called Claude Managed Agents took the Claude Platform team six months to ship, proving that hard infra work still requires hard infra work, AI or not. Second, Jarred Sumner rewrote a 500,000-plus line Bun codebase into Rust in 11 days using a tool called Fable, spending $165,000 in tokens. A task previously estimated at a year for a small team. Third, inside a 3,500-person company, the balance of engineering labor is shifting: prototyping is faster, but verification now costs more time than implementation, and code review and testing are increasingly delegated to AI.
A follow-up piece will compare Anthropic findings directly against OpenAI's internal practices. The real value in reading the full article is not the headline numbers but the specifics of where AI tooling broke down, where it held, and what that signals for engineering teams outside the lab.
[READ ORIGINAL →]