Policy-as-Code Governance for Generalist LLM Agents

A recent study published on arXiv (2605.20874) presents a governance-by-construction framework for enterprise agents, utilizing CUGA's modular policy-as-code layer. This system implements policy interventions at five key structural checkpoints: Intent Guard (prior to planning), Playbook (integrated within system prompts to guide reasoning), Tool Guide (at the boundaries of tool calls), among others. It integrates with generalist LLM agents to facilitate predictable, auditable, and compliance-conscious behavior in complex workflows without the need for model fine-tuning. The runtime governance architecture guarantees permitted actions, activates human oversight when necessary, and manages information exposure. The demonstration focuses on production environments where agents function independently across various tools and interfaces.