Decoupling Code Complexity from Newcomer Participation: A Causal Study of AI Coding Agent Adoption in OSS

Open-source software (OSS) projects rely on a steady stream of newcomers to remain sustainable. A significant concern is that AI coding agents—tools like Cursor or Claude Code that generate code from natural language—might crowd out these newcomers by absorbing the simple tasks they typically start with and by making the codebase harder to read. This paper provides a causal analysis of this concern by identifying 1,888 projects that adopted an AI agent (signaled by the first commit of a configuration file) and comparing them against matched non-adopting controls. To ensure causal validity, the authors restrict their main analysis to 603 projects that had at least six months of pre-adoption history, allowing for a robust difference-in-differences (DiD) estimation.

Contrary to the crowding-out hypothesis, the authors find no evidence that AI agent adoption reduces newcomer inflow. Across multiple estimators, newcomer participation remains stable or shows a slight, statistically insignificant increase. Furthermore, the authors find no negative impact on newcomer onboarding (measured by pull request acceptance and response times) or retention.

Regarding the mechanism, the authors confirm that AI adoption does increase code complexity, but the effect is much smaller than previously reported. Cyclomatic complexity increases by 3–4% across all languages, and cognitive complexity increases by approximately 11% in Python. Crucially, when the authors perform a unit-for-unit analysis on the specific repositories where complexity increased, they find that newcomer participation does not decline. This demonstrates that while the complexity barrier is real, it is decoupled from human participation outcomes in established projects.

This study provides a critical, evidence-based perspective on the impact of AI in software engineering. By demonstrating that the feared trade-off between AI assistance and human participation does not materialize in established projects, the authors offer reassurance to maintainers. The findings suggest that the complexity cost of AI agents is manageable and that the primary risks to newcomer pipelines are likely social or procedural rather than technical. The paper also provides a methodological template for future studies, emphasizing the necessity of validating parallel trends and using unit-specific comparisons to avoid misleading results.