Researchers released guardscope, a tool that measures how well runtime controls watch LLM agents’ actions. The framework first predicts which of an agent’s 15 possible action paths a control can see, then runs a workload on one of eight deployment topologies to label each action as checked, bypassed or prevented. Experiments used two agent harnesses and Microsoft’s governance toolkit and agentgateway. The code and framework are publicly available.
Why it matters
Enterprises can now confirm that their safety controls actually see the actions they intend to monitor.
