ETH Zurich Study: LLM-Generated AGENTS.md Files Reduce Task Success by 3% and Increase Inference Costs 20%+ Across 138 Repos

InfoQ·high signal

First comprehensive study on context files tested 5,694 PRs across 138 repositories with Sonnet-4.5, GPT-5.2, and Qwen3-30B. LLM-generated context files hurt performance by 3% and added 2-4 extra reasoning steps. Human-curated files gained only 4% while incurring the same token overhead. The recommendation: keep context files under 60 lines, limit to non-inferable details like custom build commands, and never auto-generate them with an LLM.