The modern software development lifecycle is currently defined by a relentless pursuit of velocity. For years, the pull request has been the primary bottleneck in this pipeline, where critical features and urgent bug fixes often sit in a state of limbo, waiting for a human reviewer to find a window of availability. To break this deadlock, engineering teams have rushed to integrate Large Language Models into their CI/CD pipelines, moving from simple automated linting to sophisticated AI agents that can navigate repositories and suggest architectural changes. The promise is a world where code is reviewed instantly and merged without friction, but the reality of this transition is far more complex than a simple increase in speed.
The Data Behind the AI Transition
To understand the actual impact of this shift, researchers tracked the evolution of the review process across 207 open-source GitHub projects, analyzing a massive dataset of 1.02 million completed pull requests between May 2022 and February 2026. The study categorizes the evolution of code reviews into three distinct eras: the human-centric review period, the LLM-assisted review period, and the agentic review period. The critical differentiator in this timeline is the introduction of AI agents capable of autonomous repository exploration and tool execution, moving beyond the static prompt-response nature of early LLMs.
Project adoption of these tools followed three primary trajectories. The largest group, representing 46% of the projects, followed a gradual adoption path, starting with limited LLM usage before expanding into agentic workflows. Another 32% of projects adopted an agent-focused approach, rapidly scaling their AI integration only after agentic capabilities became available. The remaining 22% were early LLM adopters, applying AI to the majority of their pull requests from the very beginning of the LLM era.
To measure whether this speed came at the cost of quality, the researchers tracked three specific anti-patterns. The first is Review Buddies, where a project relies repetitively on a narrow combination of reviewers, stifling diverse perspectives. The second is the Sleeping Review, where a pull request stagnates for an extended period without progress. The third is the Large Changeset, where an excessive amount of code is submitted in a single request, making thorough review nearly impossible. By tracking these metrics, the study provides a statistical map of how AI changes the fundamental behavior of developers.
The Paradox of Speed and Quality
The data reveals a striking divergence in how AI affects velocity depending on how it was introduced. Projects that adopted AI gradually or focused specifically on agentic workflows saw a significant decrease in the time required to reach a review decision compared to the human-centric era. However, the early LLM-focused group saw no such improvement. In these projects, the speed of review remained stagnant across both the LLM and agentic eras. The analysis suggests that these teams fell into a trap of over-reliance on specific LLM reviewer accounts, which actually reduced the diversity of perspectives and created a new kind of bottleneck.
Efficiency peaks when the collaboration sequence is optimized. The most successful pattern emerged when agents performed the initial review and humans responded to those findings, or when multiple agents collaborated before a human intervened. In 75% to 95% of these high-velocity cases, the agent acted as the first line of defense, scanning the changes and providing a concise summary. This shifted the human's role from a tedious discovery phase to a high-level decision-making phase, drastically reducing the cognitive load required to close a PR.
Yet, this acceleration comes with a measurable decline in review health. The prevalence of anti-patterns rose sharply with AI participation. While human-only reviews exhibited anti-pattern rates between 69% and 76%, reviews involving AI saw these rates climb to between 78% and 94%. The most alarming spike occurred within the Review Buddies pattern. In human-only environments, this occurred in roughly 16% of cases. With LLM participation, that number jumped to an average of 60%, and it remained high at 53% with agentic participation. This suggests that many teams are simply assigning a single AI bot to every PR, effectively replacing a diverse group of human peers with a single, consistent algorithmic perspective.
Despite the introduction of AI, certain structural constraints remain immutable. The physical characteristics of a pull request—such as the number of commits, the total size of the changes, and the volume of inline discussions—continue to be the primary predictors of a Sleeping Review. Furthermore, the social hierarchy of the repository persists; pull requests submitted by highly experienced authors continue to be reviewed faster than those from novices, regardless of whether an AI agent is involved in the process.
The evidence suggests that the blind application of AI to every pull request is a recipe for technical debt. The goal should not be the total replacement of human review, but a context-aware deployment strategy. Simple maintenance tasks and routine updates are ideal for lightweight LLM summaries and automated checks. Changes that require a deep understanding of the repository's global state are best handled by AI agents. However, core logic changes that carry high risk must remain under a human-led review system to prevent the Review Buddies phenomenon from compromising the codebase.
For engineering leaders, the lesson is that single-metric success—such as reducing the average time to merge—is a dangerous KPI. True performance must be measured by balancing review velocity against the rate of anti-patterns. By designing the AI's role based on the type of change, the size of the PR, and the experience of the author, teams can leverage the speed of agentic workflows without sacrificing the critical diversity of thought that keeps software robust.
