Longitudinal Study of Human-AI Collaboration in Open-Source Coding

A recent preprint available on arXiv (2608.13884) offers a longitudinal examination of human-AI collaboration within open-source software development, concentrating on two AI infrastructure projects: vLLM and SGLang. The analysis spans from February 2023 to June 2026 for vLLM and from January 2024 to June 2026 for SGLang, encompassing a total of 33,228 merged pull requests (18,290 from vLLM and 14,938 from SGLang). Researchers calculated seven engineering metrics, including pull request throughput and contributor diversity. The development timeline was divided into four distinct eras, reflecting significant shifts in AI-assisted software engineering. Results show marked improvements in development speed, with one project's PR throughput increasing by 21 times, suggesting potential benefits for biomedical AI agents and bioinformatics development. The study provides valuable empirical insights at the team level, an area previously underexplored. The authors remain unnamed in the abstract, but the work is categorized under cross-listing on arXiv.