In today's programming environment, coding agents can autonomously generate complete pull requests, leading to significant disagreements among practitioners regarding the role of code review. The debate centers on whether code review becomes a bottleneck, whether human review remains necessary, and whether this evolution quietly undermines human understanding of the code.
Observational analysis of public GitHub activity reveals that agent-authored pull requests are reviewed less frequently, merged faster, and discussed less than those authored by humans. However, the direction of these trends can flip under different analytical choices, indicating that while the traces of change are clear, the reasons are not.
To uncover these mechanisms, we synthesize practitioner discourse at scale into an explanatory theory: we collected 38,709 grey-literature documents (engineering blogs and Reddit threads), filtered for those substantively related to code review, and coded a stratified random sample of 3,100 using an LLM-assisted pipeline, building a causal model with 26 constructs and 67 relationships (64 directed, 3 contested).
The core claim of this theory is that the review process is the control point through which the coding agent's impact on software is determined, and AI does not simply change the sign of that effect: the direction is set by the team through the expertise its humans bring and the structure of the review process. This theory makes competing positions explicit and turns "AI is changing code review" into falsifiable propositions with defined constructs and moderators.
Additionally, we offer the underlying LLM-assisted grey-literature theory-building method as a scalable template for software engineering research, with a public implementation.
Blogger's Review: This article profoundly explores the role of AI in code review, particularly revealing potential causal relationships through the voices of practitioners. While the introduction of AI may simplify certain processes, the expertise of humans and the mechanisms of review remain crucial, warranting serious consideration from software engineers regarding their impacts.