Researchers have quantified something many developers probably suspected was happening: AI coding agents are now regularly reviewing code written by other AI coding agents on GitHub. A new study from Niruthiha Selvanayagam and colleagues constructs a large-scale dataset called CodAGE to examine this emerging pattern of "AI-to-AI" pull request reviews, finding 248,641 unique PRs where an AI-authored contribution received at least one AI-attributed review.

Building the Dataset

The research team linked AI-attributed PRs with AI-attributed review events from CodAGE, a public dataset of coding-agent-generated GitHub activity. The study distinguishes between two types of AI-to-AI reviews: "cross-product" reviews (where the author and reviewer use different AI products like Claude Code reviewing Copilot-authored code) and "same-product" reviews (where both author and reviewer use the same product). Of the 248,641 PRs analyzed, 45,269 received cross-product review while 208,145 received same-product review, with 4,773 receiving both types. Cross-product AI-to-AI review occurred in approximately 1.6% of identified agent-authored PRs—but despite the seemingly small percentage, absolute volumes were substantial and grew dramatically over time.

Explosive Growth in Bot-on-Bot Review

Perhaps the most striking finding is temporal: cross-product AI-to-AI review volume increased by more than two orders of magnitude from Q1 2025 to Q3 2025. That's a 100x-plus jump in roughly six months, suggesting that as AI coding agents proliferate across development teams, they're increasingly bumping into each other's work on shared codebases. The researchers note this closed-loop dynamic—where one AI reviews another agent's contribution—is becoming structurally embedded in how some organizations run their CI/CD pipelines.

Comment Patterns Reveal Product Personalities

The study also analyzed reviewer output quality and quantity across different author-reviewer configurations. CodeRabbit, a popular AI review tool, labeled 35.0% of its comments on Claude Code-authored PRs as refactor suggestions, compared with just 10.5% on Copilot-authored PRs. The researchers caution this difference may reflect characteristics of the PRs themselves rather than inherent reviewer behavior patterns, but the variance suggests different AI tools have distinct stylistic tendencies when reviewing code they didn't author. For three of four dual-role reviewers studied, mean comments per PR were 58-65% higher in same-product configurations—though effect sizes were small overall and concentrated in the upper tail of distribution.

Latency Metrics Show Near-Instant Bot Responses

Among review pairs with complete timestamp data, median response latency was remarkably fast: 1.2 minutes for cross-product pairs versus 4.7 minutes for same-product pairs. The researchers note differential timestamp availability and reviewer composition limit direct comparison, but the numbers suggest AI reviewers can turn around feedback in minutes rather than the hours or days human developers typically take. This speed advantage could reshape code review workflows if organizations lean into automated agent-to-agent quality gates.

Key Takeaways

  • Cross-product AI-to-AI reviews grew 100x+ from Q1 to Q3 2025, indicating this is a structural trend not a fluke
  • Same-product review (Claude reviewing Claude Code PRs, for example) dominates at roughly 4.6x the volume of cross-product
  • Comment style varies significantly by product pairing—CodeRabbit suggests far more refactors on Claude-authored code than Copilot-authored code
  • AI-to-AI review latency under 5 minutes median could enable new automated quality gates in CI/CD pipelines

The Bottom Line

This research confirms what infrastructure engineers have been watching unfold: the software development pipeline is increasingly a conversation between autonomous agents rather than a human workflow. Whether you view this as efficiency gains or a dangerous decoupling of code quality from human intent, it's happening—and fast. Teams should start thinking about governance frameworks for AI-to-AI interactions in their repos before it becomes an afterthought.