Why Small Teams Are Leading the Way in AI Coding Agent Adoption

Post Views: 16

Small teams are the primary users of AI-driven coding agents according to a study analyzing 25,264 agentic pull requests on GitHub. Researchers Maliha Noushin Raida and Daqing Hou at Rochester Institute of Technology examined repositories with over 100 stars, focusing on interactions with AI tools such as GitHub Copilot, OpenAI Codex, and Claude Code between May and July 2025. The analysis revealed a consistent pattern where a single developer typically handles review and integration of agent-generated code. In 78.9% of cases, a solitary developer reviewed and merged agent-created pull requests, with minimal changes required. This approach accounted for nearly 90% of instances where code was accepted without modification. Despite variations in project size, group reviews remained uncommon across all scales. Small teams with one to five contributors exhibited the highest frequency of AI-assisted development, averaging 50.2 agentic pull requests per repository quarterly. This contrasted sharply with medium and large teams, which generated significantly fewer such requests. Even among small teams with high AI activity—those processing over 30 pull requests during the study period—the single-reviewer model persisted. Raida noted that increased agent usage did not correlate with more distributed review processes. The review bottleneck remained a critical constraint, as developers spent substantial time evaluating agent-generated code. The study found similar merge rates between single-reviewer and multi-reviewer workflows, with 81.2% of single-reviewer pull requests and 80.3% of multi-reviewer requests being accepted. However, the nature of the tasks differed: single-reviewer workflows primarily involved feature additions, while group reviews focused on bug fixes. Repositories generating agent pull requests at a pace matching professional developer output—25 projects in the sample—exhibited the highest volume. Most projects, however, used AI tools intermittently. The research highlighted that while AI coding agents increase productivity, human oversight remains centralized, with no evidence of widespread shifts toward collaborative review practices. The findings underscore the current limitations of AI in code review workflows, emphasizing that human judgment continues to act as the primary gatekeeper for agent-generated contributions.

“The review bottleneck remained a critical constraint, as developers spent substantial time evaluating agent-generated code.”

Key Findings

The study found that 78.9% of agentic pull requests were reviewed and merged by a single developer, with minimal changes required. Small teams (1–5 contributors) accounted for 50.2 agentic pull requests per repository quarterly, far exceeding medium and large teams. Despite high AI activity, the single-reviewer model persisted, with no evidence of distributed review processes.

Review Workflows

Single-reviewer workflows focused on feature additions, while group reviews targeted bug fixes. Merge rates were nearly identical between single-reviewer (81.2%) and multi-reviewer (80.3%) workflows, highlighting the efficiency of centralized oversight.

Project Activity

Only 25 projects matched professional developer output in generating agentic pull requests, while most used AI tools intermittently. This suggests that AI adoption remains uneven across repositories.

Implications for AI in Code Review

The research underscores that AI coding agents enhance productivity but do not replace human judgment. Human oversight remains centralized, with no shift toward collaborative review practices. This highlights the current limitations of AI in code review workflows.