# On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents

**Authors:** Jai Lal Lulla, Seyedmoein Mohsenimofidi, Matthias Galster, Jie M. Zhang, Sebastian Baltes, Christoph Treude  
**arXiv:** [2601.20404](https://arxiv.org/abs/2601.20404)  
**Date:** 2026-01-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2601.20404v1/

## Key points

- AI coding agents like Codex and Claude Code are increasingly prevalent in software development
- Lack of understanding on how repository-level configuration artifacts impact agent efficiency
- Study focuses on influence of AGENTS$.$md files on runtime and token consumption of AI coding agents in GitHub pull requests
- Findings show presence of AGENTS$.$md files leads to decrease in median runtime (by 28.64%) and reduction in token consumption (by 16.58%)
- Results highlight importance of repository-level instructions in shaping behavior, efficiency, and integration of AI coding agents within software development workflows

## Summary

, , , , 

In the realm of software development, AI coding agents like Codex and Claude Code are increasingly prevalent in autonomously contributing to software repositories. However, there is a lack of understanding regarding how repository-level configuration artifacts impact the operational efficiency of these agents. This study delves into the influence of AGENTS$.$md files on the runtime and token consumption of AI coding agents that operate within GitHub pull requests. By analyzing 10 repositories and 124 pull requests, the study compares agent performance under two conditions: with and without an AGENTS$.$md file present. The evaluation includes measuring wall-clock execution time and token usage during agent operations. The findings reveal that the existence of AGENTS$.$md files correlates with a notable decrease in median runtime (by 28.64%) and a reduction in output token consumption (by 16.58%), all while maintaining a consistent task completion behavior. These results shed light on the significance of repository-level instructions in shaping the behavior, efficiency, and integration of AI coding agents within software development workflows. Authored by Jai Lal Lulla, Seyedmoein Mohsenimofidi, Matthias Galster, Jie M. Zhang, Sebastian Baltes, and Christoph Treude, this research provides valuable insights for configuring and deploying AI coding agents effectively in practical settings. Additionally, it sets forth a comprehensive research agenda to further explore the role of repository-level artifacts in optimizing the performance of AI coding agents within software development environments.