AgenticFlict merge-conflict dataset: a selected textual baseline

Primary source: Ogenrwot and Businge, AgenticFlict, April 2026. From 932,791 agent PR records as of January 5, researchers selected 142,652 open or closed-but-unmerged candidate PRs; 107,026 could be Git merge simulated; they label 29,609 as textually conflicting and report 27.67% of simulated set. This is not the fraction of production merged PRs conflicting, not the fraction of parallel agent teammates colliding, and not a causal agent-vs-human comparison. It measures git-level file conflicts, not clean-merge behavioral incompatibilities. Missing/inaccessible commit histories account for 35,626 exclusions.

Data-quality caution. Table 1 reports 29,609 conflicts plus 75,924 clean = 105,533, 1,493 fewer than 107,026 simulations, with no reconciliation visible in the manuscript. Table 2 per-agent rows total 107,026 PRs and 29,609 conflicts; adjacent prose gives slightly different percentages than its own tabulation. Accordingly treat 27.67% as the authors’ reported selected-sample statistic, not a precise deployment base rate or between-agent ranking. A matched human PR cohort at the same workflow stage, base freshness and size is needed to isolate agent contribution. STALE tests a distinct semantic failure after a clean merge; coordination brief.