From Industry Claims to Empirical Reality: Code Review Agents Studied Across 400K+ Agent-Generated PRs
arXiv·high signal
Empirical study challenging industry claims that code review agents (CRAs) can manage 80% of PRs without human involvement. With OpenAI Codex alone generating over 400,000 pull requests in two months, CRAs have become routine gatekeepers in development workflows. The study systematically measures actual CRA effectiveness against vendor claims as agentic PR volumes surge, providing the first large-scale empirical reality check on automated code review.