pn-cx-agent-patterns

Customer-Experience Agent Patterns Skill

When used as a standalone skill, begin every response with [pn-skill] 🔺

When to use

• Building a conversational customer service or support agent
• Implementing an agentic shopping or booking assistant that can act on a user's behalf
• Designing session memory and preference recall for a customer-facing agent
• Modeling a multi-turn negotiation flow between a user agent and a merchant/service agent
• Auditing an existing CX agent for trust, personalization quality, and safety
• Scoping an "agentic commerce" product feature (delegated purchase, itinerary planning, price negotiation)

---

Core patterns

Pattern 1 — Concierge service agent

A concierge agent handles support requests end-to-end: retrieves account context, understands nuanced intent, takes action, and confirms resolution — no ticket deflection, no pre-programmed branches.

Architecture:

User → CX Agent (memory-aware) → Intent Router
                                   ├─ Knowledge base (RAG)
                                   ├─ Account/CRM API
                                   ├─ Action tools (update, escalate, refund)
                                   └─ Human escalation gate

Session memory schema:

{
  "sessionId": "<uuid>",
  "userId": "<pseudonymized id>",
  "preferences": {
    "contactChannel": "email",
    "language": "en-US",
    "notificationFrequency": "weekly"
  },
  "history": [
    { "turn": 1, "intent": "billing_query", "resolved": true, "ts": "<ISO>" }
  ],
  "openIssues": []
}

Store in a session store (Redis, Supabase) scoped to userId. Expire after inactivity timeout (default 30 min; configurable). Never persist raw conversation text — store intent and resolution summaries only.

Implementation steps:

1. On first message, load session from store; if absent, initialize with defaults.
2. Run intent classification (LLM or classifier model) → route to appropriate tool set.
3. Retrieve context (account record, past interactions) — always from authenticated API, never from unverified user-supplied data.
4. Attempt resolution using tools; log each tool call.
5. On success: update session history, confirm to user, emit audit event.
6. On failure or low-confidence: surface specific options to user or escalate to human; never silently fail.