Researchers from ETH Zurich published a paper evaluating whether AGENTS.md and repository-level context files actually help coding agents. The study by Thibaud Gloaguen, Niels Mündler, Mark Müller, Veselin Raychev, and Martin Vechev examines a practice widely encouraged despite lacking rigorous investigation into effectiveness.

The team tested multiple coding agents across two settings: established SWE-bench tasks with LLM-generated context files, and real repositories with developer-committed files. Results showed that context files reduce task success rates while increasing inference costs by over 20%. These files encourage broader exploration but contain unnecessary requirements that make tasks harder.

The researchers recommend that human-written context files should describe only minimal requirements. Their findings challenge conventional wisdom that more context improves coding agent performance, suggesting developers should be more selective about what information they include in AGENTS.md and similar files.