An Empirical Study: AI Agent Rules Need Context and Layered Enforcement

# Summary A study of 2,116 statements from 84 instruction files in popular AI agent repositories reveals that 64% contain enforceable policies, but most require context and layered enforcement rather than simple rules. While developers have already written substantial safety policies for their agents, the gap between natural-language requirements and what systems can actually observe and verify remains significant—single OS-level checks cannot cover policies that depend on repository structure, task progress, or prior events. The research demonstrates that effective AI agent rule enforcement demands understanding both the behavioral policies developers intend and the system-level architecture needed to observe the contextual state required to evaluate them.