AI Agent Pull Requests on GitHub: Frequency, Structure, and Merge Conflict Rates (arxiv.org)

A recent study has explored the frequency and dynamics of pull requests (PRs) generated by AI coding agents on GitHub, revealing noteworthy trends in concurrent submissions. Analyzing the AIDev-pop dataset, which encompasses 33,596 PRs across 2,807 repositories, the research found that 40.2% of repositories have AI-generated PRs submitted simultaneously. When extending the observation window to one week, this figure rose to 53.4%. Most of these concurrent submissions were made by the same agent, pointing to a high degree of intra-agent collaboration, with cross-agent interactions being remarkably rare. The findings are significant for the AI and software engineering communities as they highlight the challenges of merge conflicts, particularly between PRs from different AI agents. The study indicated that merge conflicts were 41.7% higher for cross-agent submissions compared to intra-agent, underscoring the structural complexity in collaborative coding environments. With 84.4% of conflicts arising from code file modifications rather than dependency changes, the research suggests a pressing need for improved algorithms and tools to manage and resolve these conflicts efficiently as the use of AI in software development intensifies.