Research
PolicyGuide Compiles Policy Into a Workflow Graph and Lifts tau^2-bench Pass^4 From 0.42 to 0.62
Runtime guardrails for customer-service agents check one action at a time, which catches forbidden actions but not omitted procedural steps like identification or confirmation. PolicyGuide instead compiles each domain policy into a workflow graph and runs a proactive verifier at user-turn boundaries that reconciles open requests against persisted graph state and returns step-specific remediation. With a GPT-5.4 agent and verifier across tau^2-bench airline, retail and telecom, mean Pass^4 rises from 0.42 to 0.62, with the biggest jump on telecom (0.19 to 0.61), the most workflow-structured domain; the same workflows transfer to Claude Sonnet 4.6 and Gemini 2.5 Pro agents.
↳ Follow the thread