AI辅助代码审查效率提升但质量下降

Agents can compress review time without improving the review process itself.

A study of 1.02 million pull requests across 207 GitHub projects tracks the shift from human-only review to LLM-assisted and agentic review. Projects that adopted AI gradually, or moved rapidly to agents, saw review time fall by 2.5 and 4.5 days per KLOC (the code change in thousands of lines of code) in the agent era.

Projects that adopted LLM reviewers heavily and early saw no significant efficiency gain.

The pull-request interaction patterns explain part of that split.

Agent-initiated and multi-agent reviews reached decisions faster than human-only review under gradual and rapid-agent adoption. Agents often performed the initial inspection and summary, leaving the human reviewer with a shorter exchange.

But most AI-involved patterns had review smells in 78% to 94% of pull requests, versus 69% to 76% for human-only review.

Much of the difference came from repeatedly assigning the same AI reviewer identity, which narrows reviewer diversity.

– arxiv. org/abs/2607.13196

Title: "From Human-Centric to Agentic Code Review: The Impact of Different Generations of Generative AI Technology on Review Quality"