PolicyGuard: A Dialogue-Grounded Sub-Agent Verifier for Policy Adherence in LLM Agents

A new sub-agent verifier called PolicyGuard improves policy adherence in large language model agents by analyzing full conversation context, according to a paper submitted on 28 Jun 2026 [1]. The system provides actionable feedback for the agent's next turn rather than simply blocking non-compliant actions [2]. The researchers argue that prior work has treated policy adherence too narrowly as a safeguarding problem, focusing on external checks that block agent actions [2]. Real-world workflows, they contend, unfold across many turns and require explicit user confirmation, prerequisite reads, and attention to dialogue content rather than isolated argument values [2]. PolicyGuard addresses this by sharing the agent's view of the dialogue, reasoning over the policy in context, and offering conversation-specific remediation [1]. On the tau^2-BENCH airline benchmark, PolicyGuard improved the PASS4 metric by +12.0 percentage points for GPT-5.4, +6.0 percentage points for Claude Sonnet 4.6, and +12.0 percentage points for Gemini 2.5 Pro across four trials per setting [1][2]. Per-call analyses showed that PolicyGuard achieved higher policy-violation recall while blocking roughly half as often as argument-level guards [2]. The paper identifies three capabilities that prior safeguard work has often underestimated: full conversation context, self-reasoning over the policy and current dialogue, and conversation-specific remediation that guides the agent's next turn [1][2]. These capabilities allow the verifier to intervene more precisely than systems that examine only single argument values. The submission, dated 28 Jun 2026, appears on arXiv under the Computer Science and Artificial Intelligence category [1]. The work evaluates performance across three major commercial LLM vendors, suggesting broad applicability of the approach [2].