Anthropic says the jump in merged changes forced it to build a new source-control system. Separately, new tools guide humans through diffs and run review before a push. With the top model passing every blocking check in only 57.7% of attempts, builders should put their effort into review interfaces, quality gates and accountable release decisions rather than into generating more code.
From the Field Notes:
- Redesigning source control at Anthropic (2026-10-18)
- Redesigning source control at Anthropic \ Anthropic (2026-10-16)
- Anthropic says Claude produced most of Scion’s implementation while engineers set the architecture, scaling targets, and system invariants. (2026-10-16)
- Claude Code can write multi-file changes faster than you can review them, but you still have to sign off before merging. (2026-10-18)
- A separate public benchmark analysis evaluated 23 models for correctness and code-review quality, with its top model passing every blocking check in 57.7% of attempts. (2026-10-17)
- Releasing AI-generated software (8 minute read) (2026-10-15)
- OpenQodex: An open source AI code reviewer. (2026-10-19)