Design a multi-step customer support agent that resolves tickets, escalates when needed, and tracks resolution quality. Covers programmatic enforcement, tool description design, and escalation criteria.
Quick Reference
You are the lead architect for a customer support resolution agent built on the Claude Agent SDK. The agent handles three categories of customer requests: product returns, billing disputes, and account-level issues such as password resets or subscription changes. The system connects to four MCP tools: get_customer (retrieves customer profile and history), lookup_order (fetches order details by order ID or date range), process_refund (initiates a refund for a given order), and escalate_to_human (routes the conversation to a live agent with a structured summary). The business target is 80% or higher first-contact resolution rate — meaning the agent must resolve the issue without human intervention in at least 4 out of 5 conversations.
The agent runs in a production environment where it interacts with real customers via a chat widget. Conversations are typically 3 to 8 turns. The agent must identify the customer, understand the issue, take action using the appropriate tools, and confirm resolution — or escalate cleanly when the issue exceeds its authority. Common failure modes include: skipping customer identification, calling the wrong tool due to ambiguous descriptions, and escalating cases it could have resolved (or vice versa).