Deterministic Gates Prevent Silent Policy Violations in LLM Agents

Large Language Model agents that interact with external tools can inadvertently violate operational policies without indicating an error. This "silent wrong-state failure" occurs when a tool executes a command that, while syntactically valid, leads to a state forbidden by the domain's rules, such as cancelling a booking without proper verification. The agent's self-report or the tool's output often fails to expose these critical errors. A study using an airline domain benchmark revealed that a significant majority of agent failures were of this silent, policy-violating type. To address this, researchers proposed and evaluated deterministic, read-only "pre-execution gates." These gates inspect proposed tool calls and the current system state before any write action is permitted, effectively acting as a policy enforcement layer. Implementing these gates led to a substantial improvement in task success rates, particularly in tasks where the gates were actively engaged. The findings suggest that while LLMs may attempt policy-violating actions, a simple, deterministic verification step at the action boundary can reliably prevent a known class of silent failures, even with advanced models like GPT-5.2.