PolicyGuide Ensures LLM Agent Compliance Across Workflows.

Large Language Model (LLM) agents operating in customer service roles must strictly adhere to organizational policies when performing actions on behalf of users. Compliance failures can stem from either forbidden actions, such as approving an ineligible request, or the omission of required procedural steps, like identity verification. While runtime safeguards can prevent risky individual actions, they often lack the ability to guide an agent through a complete, multi-step workflow. PolicyGuide addresses this limitation by compiling each domain policy into a comprehensive workflow graph. At the boundaries of user turns, a proactive verifier is invoked, which leverages the persisted graph state to reconcile any open requests. This system then returns step-specific remediation, effectively guiding the agent along a policy-compliant path throughout the entire procedure. Evaluations across airline, retail, and telecom domains, using a GPT-5.4 agent and verifier, demonstrated significant improvements. PolicyGuide raised the mean Pass4 score from 0.42 to 0.62, with the most substantial gain observed in the telecom domain (0.19 to 0.61), which is characterized by highly structured workflows. The same workflows were also successfully transferred to Claude Sonnet 4.6 and Gemini 2.5 Pro agents. Further assessments showed the lowest observed attack-success rate under adversarial users and the strongest procedural compliance, validating its effectiveness in ensuring policy adherence.