AI Agents Submit Overlapping Code to GitHub; Cross-Agent Conflicts Spike to 41.7%

Researchers used the AIDev-pop dataset to conduct the first large-scale empirical study of concurrent pull request submission by AI coding agents. The dataset spans 33,596 pull requests across 2,807 repositories. The study had two main objectives: to measure how often multiple agent-authored pull requests overlap in time, and to characterize the merge conflicts that arise when those overlaps occur.

The team found that 40.2% of repositories contain at least one pair of co-active agent-authored pull requests when considering exact temporal overlap. Notably, these co-active pairs account for 79.4% of all pull requests generated by AI agents in the dataset, suggesting that concurrent submission is not a rare edge case but a normal part of how these agents operate. When the researchers relaxed the definition of overlap to allow a one-week collaboration window, the statistics shifted dramatically: 53.4% of repositories contained co-active pairs, and such pairs represented 95.0% of all agent-authored pull requests.

The composition of these pairs is heavily skewed toward same-agent submissions. Of the co-active pull request pairs, the vast majority were authored by the same agent. Cross-agent pairs—pull requests submitted by different agents to the same repository—accounted for only 0.5% of all co-active pairs and appeared in only 122 of the 2,807 repositories studied (approximately 4.3%). Despite their rarity, cross-agent interactions proved to be the most conflicted. The researchers replayed 747 actual three-way git merges, taking one unique co-active pair per repository, and measured textual conflicts. Cross-agent pairs had a conflict rate of 41.7% (with 95% confidence intervals that did not overlap with the intra-agent rate), compared to 19.8% for same-agent pairs.

Further analysis of the conflicted files revealed that the majority of conflicts occurred in source code files rather than dependency manifests: 84.4% of conflicted files involved source code modifications. Additionally, nearly 42% of conflicts fell into the structural category—modifications where one pull request deletes or modifies code while another adds or modifies the same section, creating semantic rather than purely textual collisions. These patterns point to a potential avenue for future tooling: detecting and flagging agent-to-agent conflicts before they are submitted or merged, and possibly providing agents with awareness of concurrent submissions to the same repository.

The paper addresses a gap in understanding how concurrent AI coding agents interact when working on shared repositories. Prior research on AI-generated code has largely treated agents in isolation, but as these tools become standard in development workflows, the question of what happens when two or more agents submit changes to the same codebase simultaneously becomes urgent. The study's use of the AIDev-pop dataset—a large empirical corpus of agent-authored pull requests—allows for the first systematic measurement of this phenomenon.

The distinction between intra-agent and cross-agent conflicts is striking. When the same AI agent submits multiple pull requests to a repository (accounting for the vast majority of co-active pairs), merge conflicts occur in 19.8% of cases. However, when two different agents submit overlapping pull requests, the conflict rate jumps to 41.7%—more than double. This gap suggests that individual agents tend to work in consistent patterns and modify overlapping code in ways that can be reconciled, whereas different agents may not have compatible approaches to the same code regions. The finding that nearly 42% of observed conflicts were structural (modify/delete or add/add operations) rather than simple textual misalignments indicates that the friction goes beyond simple line-level collision; agents are making incompatible edits to code structure itself.

No comments yet. Be the first to share your thoughts!