Research: AGENTS.md File Can Reduce AI Programming Agent Runtime by 29%, Output Tokens by 17%

According to 1M AI News, a research team from the Singapore Management University, Heidelberg University, Bamberg University, and King's College London has published a paper on arXiv, quantitatively assessing for the first time the impact of repository-level configuration file AGENTS.md on the efficiency of AI programming agents. AGENTS.md is an instruction file stored in the root directory of a code repository to guide AI agents on project architecture, build commands, coding standards, and operational constraints, similar to Anthropic's Claude Code's CLAUDE.md and GitHub Copilot's copilot-instructions.md, currently adopted by over 60,000 GitHub repositories.

The research team conducted paired experiments on 124 merged PRs (each with code changes of no more than 100 lines) across 10 repositories using OpenAI Codex (gpt-5.2-codex), under conditions with and without AGENTS.md. The results showed that the median runtime decreased from 98.57 seconds to 70.34 seconds (28.64% decrease) when AGENTS.md was present, the median output tokens decreased from 2,925 to 2,440 (16.58% decrease), with no significant difference in task completion behavior (Wilcoxon signed-rank test, p <0.05).

The researchers noted that AGENTS.md transformed agent guidance from "brief prompts" to a "version-controlled, reviewable, collaborative-maintained configuration artifact," recommending development teams to adopt it as a best practice in repositories. However, the study's limitations include testing only a single agent with OpenAI Codex, a small-scale PR sample, and a lack of comprehensive code correctness assessment.