AI coding agents are becoming increasingly common in modern software development workflows. However, producing code that simply works is not sufficient. AI-generated code must also follow the coding practices and conventions established within a repository to support maintainability, readability, and seamless integration. Despite this need, most existing evaluations of AI coding agents have focused primarily on functional correctness, while their ability to conform to project-specific coding styles remains largely unexplored. In this study, we conduct a large-scale empirical analysis of five AI coding agents (Claude Code, GitHub Copilot, Devin, Cursor, and OpenAI Codex) using the AIDev v3 dataset. We propose the Stylistic Conformance Score (SCS), a metric designed to quantify how closely AI-generated code aligns with repository conventions across four key dimensions: naming practices, code structure, error handling, and testing patterns. Our analysis of 12,297 merged pull requests across 924 repositories shows that Claude Code and GitHub Copilot demonstrate the strongest stylistic alignment, whereas OpenAI Codex exhibits notably lower conformance. The results indicate that current AI coding agents achieve only moderate stylistic consistency, with error handling representing the most difficult area of alignment. The proposed SCS framework offers a complementary perspective for assessing AI-generated code quality beyond traditional correctness-based evaluation.

Conference Title

International Conference on FutureTech 2026 (ICFT)

Conference Country

Jordan

Conference Date

Oct. 18, 2026 - Oct. 22, 2026

Conference Sponsor

IEEE