Plug in the Safety Chip: Enforcing Constraints for LLM-driven Robot Agents