THE LIVING ATLAS
Back to the discovery

Source record · 2026

Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?

Thibaud Gloaguen, Niels Mündler, Mark Müller, Veselin Raychev and Martin Vechev

Read the original Publication page
Type
Empirical coding-agent evaluation; preprint
Publication
arXiv · ETH Zurich and LogicStar.ai authors
Version used
arXiv2602.11988v3
Source date
29 September 2026
Recorded in Atlas
24 October 2026
License
Atlas paraphrase; no source figure reproduced.

Version 3 revised 29 September 2026. Python tasks; evaluated models include GPT-5.2 and GPT-5.1 mini. Not a test of this research workflow. No separate validity period has been established for the finding.

How Atlas uses this source

Keep Codex on track

Sections 4.1–4.4 and 5; Tables 2–3; Appendix B

Version 3: 20%/23% average cost increases for generated context files; no statistically significant task-resolution effect. No extrapolation to this workflow.

Open the source passage
Supports this claim:Keep instructions specific to the work