Adversarial testing, runtime guardrails, and continuous monitoring for AI agents in customer service, finance, and other production workflows