THE SAFETY ARCHITECTURE

CHAPTER SUMMARY Safety must be layered, because every individual safeguard can fail. Prompt-level guardrails are flexible instructions but the softest layer; schema-level gating limits what the model can request at all; dual-agent separation divides planning from acting; runtime approval with persistent permissions balances safety and convenience; tool-level validation and lifecycle hooks enforce rules in code. The key audit question is how many independent layers stand between a model's wish and a real-world effect. KEY TERMS REVIEW Governance: the rules and controls that keep a system safe, secure, and lawful. Defense in depth: stacking several independent safeguards. Single point of failure: one part whose failure removes all protection. Guardrail: something that keeps behavior within safe limits. System prompt: standing instructions given to the model. Schema: a formal description of the shape of a tool and its inputs. Tool gating: controlling which tools the model can request. Planner and executor: the reasoning agent and the acting agent in a dual-agent design. Approval fatigue: people approving without reading because of too many prompts. Hook: a place to attach custom code that runs automatically.