A resolution agent (Claude Agent SDK) handling high-ambiguity requests — returns, billing disputes, account issues — via custom MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Target: 80%+ first-contact resolution while knowing when to escalate.
If the agent skips get_customer and refunds the wrong account, the fix is a programmatic prerequisite gate (block lookup_order/process_refund until a verified customer ID exists) — not a stronger prompt or few-shot examples. Money = deterministic enforcement.
get_customer vs lookup_order misroute because their descriptions are minimal. First step: expand descriptions with input formats, example queries, edge cases, and boundaries — before few-shot or routing layers.
Escalating easy cases while attempting hard policy-exception cases is a decision-boundary problem. Add explicit escalation criteria + few-shot examples. Avoid self-confidence scores and sentiment analysis — neither tracks case complexity.
One message with three issues (missing item, overcharge, shipping delay) must be decomposed into distinct items, investigated in parallel with shared context, then synthesized into one unified resolution — not closed after issue #1.
On escalation, compile a structured summary (customer ID, root cause, refund amount, recommended action) for humans without the transcript. Keep exact amounts/dates in a persistent 'case facts' block so summarization can't blur them.
1.The agent occasionally refunds the wrong account because it skips identity verification. Best fix?
2.One message contains three separate issues. What prevents the agent from resolving only the first and closing the ticket?
3.The agent escalates easy cases but tries to handle hard policy-exception cases itself. Best first improvement?
4.A refund request exceeds the $500 self-service limit. What guarantees policy compliance?
5.get_customer and lookup_order keep getting misrouted because their descriptions are minimal. What's the FIRST step?
6.A customer's name matches multiple accounts. What should the agent do?
7.A customer explicitly asks to speak to a human. Best behavior?
8.The policy is silent on a competitor price-match request. What's the right move?
9.An order lookup returns 40+ fields but only ~5 are relevant to the return. What should happen?
10.Over a long conversation, exact refund amounts and dates start getting blurred as history is summarized. Best safeguard?
11.You escalate to a human agent who cannot see the transcript. What should the handoff contain?
12.Someone proposes having the agent self-rate confidence and escalate below a threshold. Why is that a weak escalation signal?
13.lookup_order returns 'no orders found' for a verified customer. How should the agent treat this?
14.A call to lookup_order fails with a timeout. What error handling lets the agent recover intelligently?
15.Why is 'strengthen the system prompt to always verify identity' insufficient for the refund-account problem?