Security teams deploying AI agents now have a way to check whether an agent will actually do its job before it goes live.