Customer-service LLM agents are increasingly tasked with acting on behalf of users, but ensuring they adhere to organizational policies remains a challenge. Compliance failures can stem from forbidden actions, like granting an ineligible change, or from omitting procedural requirements, such as user identification or confirmation. Existing runtime safeguards often intervene only on individual risky actions, lacking the ability to guide agents through complex, multi-step procedures.
To address this, researchers from Hugging Face have introduced PolicyGuide, a novel framework that compiles each domain policy into a workflow graph. At every user-turn boundary, a proactive verifier uses the persisted graph state to reconcile open requests and provide step-specific remediation, steering the agent along a policy-compliant path. This approach moves beyond action-local checks to whole-workflow guidance.
In evaluations across the τ²-bench airline, retail, and telecom domains using a GPT-5.4 agent and verifier, PolicyGuide raised the mean Pass^4 score from 0.42 to 0.62. The most significant improvement was observed in the telecom domain, which is the most workflow-structured, jumping from 0.19 to 0.61. The same workflows also transferred effectively to Claude Sonnet 4.6 and Gemini 2.5 Pro agents.
Complementary evaluations highlighted PolicyGuide's robustness: it achieved the lowest observed attack-success rate under adversarial users and demonstrated the strongest procedural compliance in an author-designed workflow-level validation. These results suggest that PolicyGuide offers a promising solution for ensuring policy adherence in customer-service LLM agents.