Is AGENTS.md useful? What the research found (gethrbr.com)

Recent research from ETH Zurich and various tech studies examined the effectiveness of AGENTS.md, a context file meant to enhance AI coding agents' performance. The findings suggest that hand-written files can slightly improve success rates for coding agents, while AI-generated files tend to have a negative impact. Notably, the cost of using context files increased by about 20%. The analysis highlights that while the structure of these context files—such as file size and organization—had minimal influence on performance, the content's specificity remains vital; developers must include essential commands and non-standard practices without overwhelming the agent with unnecessary information. Moreover, a comparison revealed that a compressed documentation index within AGENTS.md has a significant advantage over skills-based approaches, achieving a 100% pass rate on Next.js 16 APIs compared to just 53% for skills rarely invoked by the agent. The research underscores the importance of clear, concise directives tailored to each agent's tasks, and advocates for human-written documents, eliminating redundant information while focusing on crucial operational nuances. This indicates a shift towards refining how context is provided to AI agents, improving their compliance and performance in real-world coding scenarios.