Gergely Orosz

Gergely Orosz, author of The Pragmatic Engineer newsletter, published two pieces a day apart in September 2025 that approach coding agents from opposite ends: how one is built, and how anyone would know whether it worked.

The first, dated 2025-09-15, draws on interviews with Boris Cherny, Sid Bidasaria and Cat Wu about the internals of «Claude Code»: a September 2024 prototype that could only identify the music playing, until file system and bash access let it read a codebase on its own. The reported figures are specific and awkward for anyone hoping to generalise. Roughly 90% of the tool's code is written by the tool itself. Anthropic's engineers merged 67% more pull requests while headcount doubled. The team ships 60 to 100 internal builds a day.

The second, dated 2025-09-16 and co-written with Laura Tacho, surveys 18 companies including Google, GitHub, Microsoft and Dropbox and finds the measurement side unresolved. 85% of engineers use AI tools; their leaders struggle to say what that is worth. Orosz and Tacho propose the AI Measurement Framework, pairing adoption and spend with core delivery metrics like change failure rate, and warn that unmanaged AI-assisted development can produce the largest pile of technical debt yet. Acceptance rate, once the reference metric, is declining in usefulness. Orosz predicts agent telemetry is where the measurement work moves next.