A production-style AI agent safety framework demonstrating multi-layer guardrails with deterministic checks, model-based classification, policy-driven controls, risk scoring, PII protection, prompt-injection detection, human-in-the-loop approvals, and comprehensive audit logging.