Risk-based approvals, not engineer-dependent merge gates, are the future

Anthropic published their code review process last week (link in comments). No surprises that their results are mind blowing! We don't have their infinite frontier token budget but we can learn from them nevertheless.

𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰’𝘀 𝗺𝗼𝗻𝗼𝗿𝗲𝗽𝗼
2.5K changes/day in May → 30K+ in September, a 12x jump

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝘁𝗲𝗮𝗺
~90% of code is AI-written
Merged PRs >2x between June–September
Revert rate fell >40%

Low-risk changes auto merge after automated checks. Higher-risk changes require 1 or 2 engineer approval, depending on the magnitude of the risk.

𝗧𝗼𝗺𝗼𝗿𝗿𝗼𝘄: the underlying agent infrastructure that made this possible.