Claude Certified Architect - FoundationsCCAR-F Learning Guide
Case study · fictional company

Cedar Ridge Outfitters

An online outdoor-gear retailer building a customer support resolution agent on the Claude Agent SDK, with custom MCP tools for customers, orders, refunds and escalation.

The business

Cedar Ridge Outfitters, based in Denver, Colorado, sells camping, climbing and paddling gear online in North America. Its support team of 45 people handles about 52,000 contacts a month: returns, damaged deliveries, billing disputes, loyalty points and account changes.

The agent

Cedar Ridge is building a support agent with the Claude Agent SDK. It talks to customers in chat and reaches backend systems only through four custom MCP tools on a server named support:

ToolWhat it doesNotes
get_customerFinds a customer by email, phone, account number or nameName searches can return several matches
lookup_orderReturns one order by order numberReturns about 40 fields; returns use about 6 of them
process_refundIssues a refund to the original payment methodAmount in dollars and cents
escalate_to_humanOpens a ticket in the human queueHuman agents do not see the chat transcript

Targets and rules

  • First-contact resolution: 80% or better. Today's human-only figure is 64%.
  • Identity first. The customer must be verified with get_customer before any order lookup or refund.
  • Refund limit. The agent may refund up to $500 on its own. Anything larger needs a human.
  • Policy. Returns are accepted within 45 days. Price adjustments apply only to purchases on Cedar Ridge's own site; the policy says nothing about matching competitors.
  • Escalate when the customer asks for a person, when policy is silent or ambiguous, or when the agent cannot make progress.

Known data quirks

  • get_customer returns timestamps as Unix epoch seconds; lookup_order returns ISO 8601 strings; process_refund returns a numeric status (0 success, 2 pending, 5 rejected).
  • About 3% of chats mention two or more separate problems in one message.
  • Long chats (20+ turns) are where most wrong amounts and mixed-up order numbers appear in testing.

People

StakeholderCares most about
Head of Customer Care (sponsor)Resolution rate and customer satisfaction
Finance controllerNo refunds above the limit, ever
Support team leadClean handoffs her team can act on without re-asking the customer
Platform engineerA loop and hooks that are simple to reason about

Exercises set here