If you do not govern agent actions your safety boundary is at risk. Prompt injection, tool misuse, and over-broad credentials turn into real side effects fast. Nomos applies zero-trust controls at the moment an agent tries to do something real. It does not restrict the model's reasoning. It controls what the agent is actually allowed to do.