AI coding assistants boost dev speed 21x in new study
A new arXiv study of 33K+ pull requests shows AI coding tools like vLLM increased throughput 21x...
A new longitudinal study published on arXiv (arXiv:2608.13884) reveals how AI coding assistants like vLLM and SGLang have transformed open-source software development. Researchers Jiada Li and Xuesong Ye analyzed 33,228 pull requests (PRs) from two high-velocity AI infrastructure repositories—vLLM (18,290 PRs, Feb 2023–Jun 2026) and SGLang (14,938 PRs, Jan 2024–Jun 2026)—segmenting development into four eras based on AI adoption milestones.
The data shows staggering improvements in development velocity. PR throughput surged 21x in vLLM and 17.9x in SGLang, with human developers driving nearly all of this growth (bot-authored PRs accounted for <0.2%). Median cycle times plummeted to 1.04 days (vLLM) and 0.62 days (SGLang), while P90 cycle times reached 16.8 and 14.3 days, respectively. Contributor diversity also increased, with monthly unique authors rising steadily, and PR comment density grew 4.2x (vLLM) and 3.8x (SGLang), with bots contributing 15–20% of the increase. Notably, PR size remained stable, suggesting AI tools enhanced efficiency without bloating code.
- PR throughput in vLLM and SGLang increased 21x and 17.9x respectively, driven almost entirely by human developers (<0.2% from bots)
- Median cycle times dropped to 1.04 days (vLLM) and 0.62 days (SGLang), with P90 times of 16.8 and 14.3 days
- PR comment density rose 4.2x (vLLM) and 3.8x (SGLang), with bots contributing 15–20% of the increase, indicating deeper AI-human collaboration
Why It Matters
Proves AI coding assistants dramatically accelerate open-source development while preserving human oversight and collaboration.