In a nutshell: AI-agent code reviews accelerate code review decisions measurably, but do not improve review quality – a central challenge in automating quality assurance.
An analysis of over one million pull requests shows that AI-agent-based code reviews increase decision speed but do not lead to better review quality. This calls into question a central assumption in automating the code review process.
Researchers examined 1.02 million code reviews from 207 GitHub projects that evolved across three phases: purely human reviews, LLM-supported reviews, and agent-supported reviews. The study identified three adoption patterns: gradual AI adoption, rapid LLM adoption, and rapid AI-agent adoption.
The results are nuanced: agent-supported collaboration patterns – particularly when AI agents initiate reviews or multiple AI systems are involved – led to faster review decisions under both gradual and rapid agent adoption. These efficiency gains were especially evident when modeling pull request discussions as reviewer interaction sequences.
The critical finding: these speed improvements did not translate to better review quality. The analysis shows that review activity and pull request type remain consistently relevant, while human-AI collaboration patterns become the strongest explanatory factor for review efficiency once LLM and AI-agent reviewers come into play.
For engineers and engineering teams, this means: AI-supported code reviews require deliberate design to avoid trading speed for thorough quality control. The empirical findings provide concrete guidance on how human and AI-based reviewers can work together meaningfully without compromising security or code quality.
Source: arxiv.org · Published July 13, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 of the EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.