Design a multi-agent orchestration for customer service escalation

• Define a supervisor agent responsible for intent classification, confidence scoring, and routing policy enforcement • Create domain-specific worker agents (billing, fraud, returns) with scoped topics, actions, and knowledge bases • Use session context objects to pass customer ID, case history, and prior agent outputs across handoffs • Implement human-in-the-loop triggers when confidence falls below threshold or sensitive PII/fraud keywords appear • Apply guardrails at orchestration layer: action allowlists, rate limits, and audit logging for every agent decision • Design fallback paths when a worker agent fails or times out—default to human queue with summarized transcript • Measure per-agent containment rate, escalation rate, and average handle time to tune routing rules