THE SAFETY ARCHITECTURE

CHECK YOUR UNDERSTANDING • Why must layers in defense in depth be independent? ◦ Answer: If they rely on the same weakness, they fail together and give no extra protection. • Why is schema-level gating stronger than a prompt rule? ◦ Answer: The harness enforces it in code, so it does not depend on the model choosing to obey. • How does dual-agent separation resist prompt injection? ◦ Answer: Even if the planner is fooled, the executor applies independent checks and does not read the untrusted text. • What problem do persistent permissions help solve? ◦ Answer: Approval fatigue, by remembering approvals for a defined scope while still asking about new kinds of actions. • Why validate inside tools too? ◦ Answer: Tool-level validation protects against any source of bad calls and sits closest to the real-world effect.