Agentic coding systems increasingly depend on repository-level agent files such as AGENTS.md to communicate project knowledge and developer intent. These files are meant to guide automated code generation, but their effectiveness likely depends on what kinds of instructions they contain.
A new arXiv study takes up this question by examining the role of agentic instruction categories for code generation across five Python libraries. The researchers set out to determine which categories of agent instructions support efficient coding workflows, though the abstract notes that it remains unclear which specific categories matter most.
The study's significance lies in its attempt to move beyond generic advice about agent files and toward a more systematic understanding of instruction design. By focusing on real Python libraries, it aims to give developers and tool builders a clearer basis for writing AGENTS.md files that actually improve code generation outcomes.