Runtime action verification for LLM agents

How a checker stops a harmful agent action just before it runs