As autonomous coding agents become active contributors in software engineering, a key question emerges: does it matter which agent produces the code, or only what the code contains? Using the AIDev dataset (31,284 PRs from five AI agents), we develop machine learning models to predict merge outcomes at submission time, achieving 82% accuracy (ROC-AUC: 0.85). Ablation analysis reveals that PR quality characteristics (size, complexity, and content type) are collectively the most important predictors, while agent identity has negligible predictive value. Repository context emerges as the strongest standalone signal, suggesting where a PR is submitted matters as much as what it contains. Although individual reputation features rank highly in importance, they prove largely redundant with each other. These findings suggest merge outcomes for agent-authored code depend on technical merit and project context rather than agent source, with practical implications: developers deploying agents should prioritize PR quality and repository selection over agent choice.

@inproceedings{Selvanayagam26,
  title = {Does the Agent Matter? Predicting Merge Outcomes of AI-Authored Code},
  author = {Niruthiha Selvanayagam},
  booktitle = {FSE},
  year = {2026},
  doi = {10.1145/3803437.3804439}
}