Type: url
Original: https://arxiv.org/pdf/2601.20404
Status: generated
Summary
“On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents” (Lulla et al.; ICSE JAWs 2026) runs OpenAI Codex on 124 small pull-request tasks from 10 GitHub repositories, each twice: once with the repository’s root AGENTS.md file and once with it removed. With AGENTS.md present, median wall-clock runtime fell by 28.64% and median output token use by 16.58%, with comparable task completion behaviour in a manual sanity check. The authors frame this as initial evidence that repository-level instruction files have measurable operational effects, and outline a research roadmap extending the study to other agents, larger tasks and output correctness.