Study Evaluates Impact of AGENTS.md on AI Programming Efficiency

A recent study conducted by research teams from Singapore Management University, Heidelberg University, Bamberg University, and King's College London has been published on arXiv, assessing the impact of AGENTS.md on AI programming agent efficiency. According to BlockBeats, AGENTS.md is an instruction file located in the root directory of code repositories, designed to guide AI agents on project architecture, build commands, coding standards, and operational constraints. It is similar to Anthropic Claude Code's CLAUDE.md and GitHub Copilot's copilot-instructions.md, and has been adopted by over 60,000 GitHub repositories. The research involved paired experiments using OpenAI Codex (gpt-5.2-codex) on 124 merged pull requests across 10 repositories, each with code changes not exceeding 100 lines. The experiments were conducted under conditions with and without AGENTS.md. Results indicated that the presence of AGENTS.md reduced the median runtime from 98.57 seconds to 70.34 seconds, a decrease of 28.64%, and reduced the median output tokens from 2,925 to 2,440, a decrease of 16.58%. There was no significant difference in task completion behavior (Wilcoxon signed-rank test, p < 0.05). Researchers highlighted that AGENTS.md transforms agent guidance from "transient prompts" to "version-controlled, reviewable, collaboratively maintained configuration artifacts," recommending its adoption as a standard practice in repositories. However, the study's limitations include testing only a single agent, OpenAI Codex, focusing on small-scale pull requests, and lacking comprehensive code correctness evaluation.