Production-readiness review and hardening

You have an agent that works in a demo. This project asks the question a supervising partner, a security reviewer and an on-call engineer would each ask before it touches a real matter: what happens when it's attacked, when it crashes, when it's audited, when it gets expensive, and when the model changes? You'll write a threat model, then make one fix per area, each with a test that proves it works.