Controlled StudyAcross 1.02 million reviewed pull requests, agent-involved review patterns are associated with faster review decisions under Gradual AI Adoption and Rapid AI Agent Adoption, but the efficiency gains do not translate into fewer review smells — agent-init and multi-agent reviews carry review smells more often than human-only reviews, and Rapid LLM Adoption shows lower quality with no efficiency gain.
From Human-Centric to Agentic Code Review: The Impact of Different Generations of Generative AI Technology on Review QualityWhat it means Adopt AI reviewers for latency, not assurance — keep quality-bearing checks (tests, security scans, human sign-off on risky paths) independent of the AI review lane.
arXiv preprint · JUL 14
Controlled StudyDelivering the same requested code changes as structured line-anchored comments instead of a holistic prompt cut generated tokens by 22% (Claude Opus) to 58% (Claude Sonnet) — 24-80% on files of 100+ lines — while correctness rose 2 points pooled and 5-7 points for three of five local models.
Line-Anchored Feedback Cuts Token Costs and Improves Correctness in AI Code EditingWhat it means Anchor revision feedback to specific lines rather than describing changes in prose — it's a cheap format change that cuts token spend and helps smaller models most.
arXiv preprint · JUL 14
Large Benchmark38.9% of agent-generated pull requests contain at least one security smell (82.3% of smells are supply-chain integrity issues), but humans introduced 67.6% of the genuine leaked secrets in these agent-assisted workflows, and existing review processes missed 81.1% of credentials before integration.
Trust but Verify? Uncovering the Security Debt of Autonomous Coding AgentsWhat it means Put secret-scanning and supply-chain checks at the point of human-AI collaboration, not just on the agent's output — the human side of the workflow is leaking most of the credentials, and review is catching almost none of them.
arXiv preprint · JUL 14