This research from folks at ETH Zurich challenges the emerging practice of keeping repo context on AGENTS.md 🤔 (at least for python code)

context files tend to reduce task success rates compared to providing no repository context, while also increasing inference cost by over 20%

our results suggest that context files have only marginal effect on agent behavior, and are likely only desirable when manually written

https://arxiv.org/abs/2602.11988

Very interesting read, my assumption was context files decrease token consumption since there is less for the LLM to explore thinky spin

Just linking this other discussion here, for visibility https://openhands-ai.slack.com/archives/C06R25BT5B2/p1771278117634859?thread_ts=1771277791.461639&cid=C06R25BT5B2