New Research Reassesses the Value of AGENTS.md Files for AI Coding
Summary
A ETH Zurich study challenges the value of AGENTS.md-style context files for AI coding. Using AGENTbench with 138 Python tasks and four agents across three scenarios (no context, LL-generated context, human-written context), the results show that LL-generated context increases effort and costs while reducing task success slightly, whereas human-written context yields only marginal gains at higher costs. The authors argue context files offer limited value and should be carefully crafted by humans, revealing a gap between current guidance and observed outcomes.