This study sets out to characterize agentic pull requests — those authored by AI coding agents — in comparison with human-generated pull requests, and to examine how their properties change across different stages of the software development lifecycle. The authors’ premise is that while developers increasingly benefit from coding agents, the agents’ impact on software quality remains insufficiently understood, and that how agentic contributions evolve across the lifecycle has not been thoroughly investigated.
The analysis is built on the AIDev dataset. Working from it, the authors first analyze how differences in merge rates between agentic and human-generated pull requests vary over time. They then identify the types of development task to which AI coding agents are predominantly applied, and investigate how those task distributions evolve across development quarters. Finally they compare a set of key characteristics of agentic and human-generated pull requests, with attention to what those characteristics imply for software quality and to how they change over time.
The authors present the contribution as an empirical and longitudinal perspective on the role of AI coding agents in software development, offering what they describe as a more nuanced understanding of the benefits and limitations of these agents in real-world practice.
Key Points
- Characterizes agentic pull requests against human-generated ones and tracks how their properties change across stages of the software development lifecycle
- Uses the AIDev dataset as its material
- Analyzes how merge-rate differences between agentic and human-generated pull requests vary over time
- Identifies the development task types AI coding agents are predominantly applied to, and how those task distributions evolve across development quarters
- Compares key characteristics of the two kinds of pull request with respect to their implications for software quality
Notes
The paper is positioned as longitudinal — its distinguishing move is tracking change over development quarters rather than taking a single snapshot. Its authors present the contribution as a more nuanced understanding of the benefits and limitations of coding agents in real-world practice, rather than as a verdict on their impact on software quality.