Type: url
Original: https://arxiv.org/pdf/2605.02273
Status: generated
Summary
This EASE 2026 short paper analyses code review interactions on AI-generated pull requests in GitHub repositories drawn from the AIDev dataset, comparing them with human-authored pull requests in the same repositories. It reports that most agent-authored PRs receive no recorded review and that, when reviewed, the discussion is dominated by agents, with human participation frequently expressed as agent-steering commands rather than standalone evaluation. The authors conclude that conventional review metrics should be interpreted cautiously as indicators of human oversight in agentic workflows.