Human-AI Synergy in Agentic Code Review

AI agents are increasingly deployed in code review workflows, yet their role relative to human reviewers remains poorly understood. This large-scale empirical study analyzes 278,790 code review conversations across 300 open-source GitHub projects to compare AI agents and human reviewers in collaborative agentic code review.

Key findings:

• Human reviewers provide broader feedback covering understanding, testing, and knowledge transfer — areas AI agents currently lack
• Reviewers exchange 11.8% more rounds when reviewing AI-generated code than human-written code
• AI agent suggestions are adopted into codebases at significantly lower rates than human suggestions
• Over half of unadopted AI suggestions are either incorrect or resolved through alternative developer fixes
• When AI suggestions are adopted, they produce significantly larger increases in code complexity and code size

While AI agents are effective for scaling defect screening, human oversight remains essential for suggestion quality and contextual feedback. This study informs the design of future human-AI collaborative agentic code review systems.

Authors: Suzhen Zhong, Shayan Noei, Ying Zou, Bram Adams
arXiv: 2603.15911