Study reveals cautious adoption of AI coding agents across GitHub projects

A large-scale analysis of how GitHub projects integrate AI coding agents reveals uneven adoption patterns and human oversight challenges across the open-source ecosystem.

Researchers examined over 25,000 pull requests generated by agentic coding tools across 2,361 popular repositories to understand how these autonomous systems are being deployed in real-world development environments. The study uncovers significant gaps between early-adopter enthusiasm and mainstream use of AI-assisted code generation.

Adoption remains concentrated#

The findings indicate that mainstream adoption is far from universal. The median repository generated only one to two agentic PRs over a three-month period, suggesting that most projects remain cautious or inactive users of these tools. However, adoption does vary by project size: smaller codebases with just 1–5 contributors showed proportionally higher engagement with agentic systems than medium and large projects, possibly because smaller teams have fewer barriers to experimentation.

Productivity gains are uneven#

Project-level productivity from agentic contributions shows substantial variation. While a minority of projects exceeded an industry benchmark of 36 pull requests per contributor quarterly, most remained well below that threshold. This disparity hints that effective use of AI coding agents may depend heavily on how well they align with individual project workflows and team structures.

Human-agent collaboration remains simplistic#

Perhaps most notably, human oversight of agent-generated work follows a remarkably narrow pattern: a single developer typically reviews and merges AI contributions, with multi-developer collaboration on agent PRs remaining rare. This concentration of oversight in one person per project raises questions about scalability, knowledge sharing, and the long-term sustainability of human-AI workflows in software teams.

What comes next#

The researchers characterize their work as an early snapshot of agentic tool adoption during a formative period. They argue that the success of AI coding agents will ultimately depend not only on technical capability improvements but also on how teams design organizational processes and governance around autonomous contributions.