The customer support resolution agent is the scenario most people picture when they think of Claude in production: a customer writes in, and an agent looks up orders, checks accounts, issues refunds or hands the case to a person. It touches three domains at once — agentic architecture, tool design and context management — which is why questions set here tend to test how the pieces fit together rather than any single feature.
The system you should picture
An application receives the customer's message, attaches the authenticated customer identity from its own session, and calls the Messages API with a system prompt (role, policies, tone, escalation rules) and a handful of tools: look up an order, read account status, issue a refund, create a ticket for a human. Your code runs each requested tool, returns tool results, and loops until Claude finishes or a limit is reached. A human support queue sits behind the agent.
Decisions the exam probes
Where rules are enforced. A refund threshold, an identity check or a "never reveal another customer's data" rule must hold every time, so it belongs in the tool handler or the application layer, not only in the system prompt. Expect at least one option that relies on wording alone.
Loop limits and failure. Malformed records, timeouts and unexpected inputs make agents loop. A maximum number of turns or a cost budget, with escalation when it is reached, turns a runaway into a handled case. Tool failures should come back as clear error results that say whether a retry is safe.
When to escalate. Explicit requests for a person, repeated contacts, policy thresholds and out-of-scope questions are the classic triggers. A good handoff is a structured summary — issue, actions taken, what the customer wants, why the agent escalated — with the transcript attached.
Cost and latency. The same long system prompt and tool definitions go out with every request, which makes prompt caching the obvious lever, provided the cached prefix stays identical. Long conversations accumulate stale tool output that should be summarised or trimmed.
Common traps
- Letting the model supply the customer ID used for lookups.
- Returning an empty result when a tool fails, so the agent concludes there is no data.
- Fixing runaway loops with a prompt instruction instead of a counter.
- Putting a timestamp or request ID at the top of the system prompt, which breaks caching.
- Escalating with a one-line note, which forces the customer to repeat everything.
How to practise
Build a toy version with two tools and a fake order database. Make one tool time out, give one customer a malformed record, and add a refund rule with a threshold. Then watch what happens with and without an iteration cap and a handler-level policy check. Ten minutes of breaking this system teaches more than an hour of reading about it.
The six scenarios are listed in Anthropic's official exam guide (version 1.1), with four drawn per sitting. The emphasis described here is our own study guidance, not official exam content.