AGENTS.md (or GEMINI.md) files are injected into every coding agent session, making them high-leverage but also high-risk. Research from ETH Zurich shows auto-generated AGENTS.md files reduce task success rates by ~3% while increasing inference costs by 20%+. Human-written files only marginally improve performance (~4%) and still raise costs. Key findings: codebase overviews don't help agents navigate faster, tools mentioned in AGENTS.md get used 160x more often, and agents can follow ~150–200 instructions but the harness already consumes ~50. Best practices: keep it under 300 lines (ideally under 60), include tech stack, project purpose, and non-obvious build/test tooling. Avoid directory listings, code style rules, task-specific instructions, and auto-generated content. Use progressive disclosure by pointing to separate task-specific docs rather than embedding everything.
