AgenticFlict: 27.67% of 107K unmerged agent PRs no longer merge cleanly, from 15% for Copilot to 32% for Codex
Ogenrwot and Businge replayed git merges for 107,026 open or closed-unmerged agent PRs from the AIDev dataset and found 29,609 with textual conflicts and 336,380 conflict regions, with conflict rates varying sharply by agent.
Source: AgenticFlict: A Large-Scale Dataset of Merge Conflicts in AI Coding Agent Pull Requests on GitHub
Most talk about agent PR throughput is about review: who reads the diff, and how fast. AgenticFlict, by Daniel Ogenrwot and John Businge (UNLV, AIware '26), measures a different place work stalls: the PR can no longer merge into its base branch.
They took 142,652 agent PRs from 59,412 repositories in the AIDev dataset and, for the open and closed-but-unmerged ones, ran a deterministic local merge (git merge --no-commit --no-ff of the PR head into its base). Of 107,026 PRs they could simulate, 29,609 conflicted, a rate of 27.67%, producing 336,380 conflict regions with file and line spans. The paper calls the conflicts "both frequent and often substantial."
The spread across agents is wide (Table 2, with 95% confidence intervals):
- Copilot: 15.24% of 16,954 PRs
- Cursor: 19.75% of 7,196
- Devin: 22.85% of 8,241
- Claude Code: 25.93% of 779
- OpenAI Codex: 31.85% of 73,856
Two cautions before reading too much into those. The sample is PRs that were not merged, so this is a conflict rate among stalled or abandoned work, not among everything agents open. And the authors call their analysis preliminary and exploratory; a conflict in a PR that never merged can be a cause of the stall or a result of it sitting. The Claude Code sample is small. Dataset and code are on Zenodo.
Why it matters for code review
When agents open PRs around the clock, the base branch keeps moving while each PR waits for review. Every hour of queue time makes a textual conflict more likely, and the rework lands on whoever picks it up. AgenticFlict gives that cost a number: more than a quarter of the agent PRs that didn't merge also no longer apply. That makes review latency, rebase policy and small diffs part of the same throughput problem, not separate ones.