Agentic AI
AI Agents' Concurrent GitHub Pull Requests Often Collide
A new study analyzing tens of thousands of AI-generated pull requests reveals how often multiple agents submit changes to the same repository at overlapping times—and how frequently their edits conflict.
When multiple AI coding agents target the same repository, they frequently end up submitting pull requests simultaneously—creating merge headaches that vary dramatically depending on whether the agents are the same or different.
The Scale of Concurrent Submissions#
Researchers examining the AIDev-pop dataset, which spans 33,596 pull requests across 2,807 repositories, found that agent-generated PRs overlap far more often than might be expected. Using exact temporal overlap as a measure, 40.2% of repositories contained at least one pair of simultaneously active agent-authored pull requests, and those overlapping pairs accounted for 79.4% of all agent-generated PRs. When the researchers widened their window to one week of collaboration activity, the numbers jumped dramatically to 53.4% of repositories and 95.0% of all PRs.
The overwhelming majority of these concurrent submissions came from the same agent working on multiple changes. Cross-agent collaboration conflicts—where two different agents submitted overlapping PRs—occurred in only 122 out of the 2,807 repositories studied (roughly 4.3%), representing just 0.5% of all co-active PR pairs.
Merge Conflicts Hit Harder with Cross-Agent Work#
The real problem emerges when different agents collide. In a three-way git merge simulation across 747 unique concurrent PR pairs, researchers computed how often actual textual conflicts arose during the merge process.
Cross-agent pairs experienced significantly more merge conflicts than same-agent pairs: 41.7% of cross-agent merges generated textual conflicts, versus only 19.8% for intra-agent pairs. The researchers confirmed this gap was statistically significant with non-overlapping 95% confidence intervals.
The Nature of the Conflicts#
Most conflicts centered on source code files rather than dependency manifests—84.4% of conflicted files were source code. Nearly 42% of the observed conflicts fell into the structural category, meaning they involved modify/delete or add/add operations that go beyond simple line-based text collisions. This suggests the conflicts reflect deeper divergence in how different agents approach the same codebase, rather than minor overlaps in adjacent lines.
