FEFreeExamDumps.in

CCA-F Practice Questions — Page 13

Question 121

Open question ↗

Scenario: Customer Support Agent

A support agent must never auto-issue a refund above $1,000 without human approval. The team added "Never refund more than $1,000; escalate larger amounts" to the system prompt, yet a quarterly audit finds that roughly 3% of refunds over the limit were still auto-approved.

What is the most reliable way to enforce this limit?

  • A.Rewrite the system-prompt rule more forcefully and add a few-shot example of declining a $1,500 refund.
  • B.Lower the model temperature to 0 so it follows the instruction deterministically.
  • C.Add a `PreToolUse` hook on `process_refund` that redirects any amount over $1,000 to `escalate_to_human`.
  • D.Move the refund limit into the project `CLAUDE.md` so it loads on every turn.

Question 122

Open question ↗

Scenario: Customer Support Agent

The team is tuning when the agent should escalate to a human. Designers propose three signals: customer sentiment analysis, a model self-rated 1–10 confidence score, and explicit rules. Logs show sentiment flags many simple FAQ tickets and the confidence score is poorly calibrated.

Which basis for escalation is most reliable?

  • A.Explicit rule-based triggers: customer asks for a human, policy is silent on the request, or the agent cannot make progress — reinforced with few-shot examples.
  • B.Sentiment analysis, because frustrated customers always need a human.
  • C.The model's self-rated confidence, auto-escalating whenever it falls below 6/10.
  • D.A custom classifier trained on historical tickets, regardless of label availability.

Question 123

Open question ↗

Scenario: Customer Support Agent

A customer says "I want a refund on my recent order." The `get_customer` tool returns three accounts matching the same name and email domain. The agent currently picks the most recently active account and proceeds with the refund.

What should the agent do instead?

  • A.Continue with the most recently active account, since it is the most likely match.
  • B.Ask the customer for an additional identifier (order number, account ID, or ZIP) to disambiguate before acting.
  • C.Process refunds on all three accounts to be safe and reconcile later.
  • D.Escalate to a human immediately without attempting any disambiguation.

Question 124

Open question ↗

Scenario: Customer Support Agent

An engineer's agent loop calls Claude with MCP tools. To decide when a ticket is resolved, the loop scans the assistant text for "all done" and also stops after `max_iterations=6`. In production it sometimes halts mid-task and other times keeps looping after the work is actually finished.

What is the correct completion signal?

  • A.Keep the text scan but raise `max_iterations` to 20 to give the agent more room.
  • B.Stop as soon as the assistant emits any text block alongside its tool calls.
  • C.Ask the model to return a boolean `is_resolved` field each turn and stop when it is true.
  • D.Continue while `stop_reason == "tool_use"` and stop when `stop_reason == "end_turn"`.

Question 125

Open question ↗

Scenario: Customer Support Agent

A customer writes: "My last order arrived damaged, I was double-charged on my card, and I also want to update my shipping address." The agent handles only the damaged-item refund and closes the ticket, leaving the billing and address issues unaddressed.

What is the correct handling pattern?

  • A.Always handle only the first issue mentioned and ask the customer to open separate tickets for the rest.
  • B.Decompose the message into three distinct items (refund, billing dispute, address change) and resolve or route each.
  • C.Merge all three into one `resolve_everything` tool call to keep the workflow simple.
  • D.Escalate the entire ticket to a human because it contains more than one request.

Question 126

Open question ↗

Scenario: Customer Support Agent

On escalation the agent currently passes the human operator one line: "Customer is unhappy, please assist." The operator has no access to the conversation transcript and must re-gather the customer ID, the order, and what the agent already tried — adding ~4 minutes per handoff.

What should the escalation handoff contain?

  • A.A self-contained structured summary: customer ID, the issue, order ID, actions already taken, and a recommended next action.
  • B.A link to the raw model logs so the operator can read the full reasoning trace.
  • C.Only the customer's most recent message, to keep the handoff short.
  • D.The agent's self-rated confidence score for the case.

Question 127

Open question ↗

Scenario: Customer Support Agent

The `lookup_order` MCP tool fails because the orders backend timed out. It returns `{"isError": true, "content": "Operation failed"}`. The agent cannot tell whether to retry, change the query, or escalate, and ends up looping on the same call.

What should the error response include?

  • A.Just a longer human-readable sentence describing what went wrong.
  • B.Nothing extra; the agent should automatically retry every `isError` response.
  • C.Structured metadata: `errorCategory` (e.g., transient), `isRetryable`, a clear message, the attempted query, and any partial results.
  • D.An empty success payload so the workflow keeps moving.

Question 128

Open question ↗

Scenario: Customer Support Agent

Compliance requires that `verify_identity` always runs and returns a verified result before `process_refund` can execute. The team debates whether to enforce this with a carefully worded system prompt or programmatically.

Which approach gives a deterministic guarantee?

  • A.A detailed system prompt explaining the required order with a worked example.
  • B.A programmatic precondition that blocks `process_refund` until `verify_identity` has returned a verified result.
  • C.Setting `tool_choice: "any"` so the model must always call a tool first.
  • D.Listing `verify_identity` before `process_refund` in the `allowedTools` array.

Question 129

Open question ↗

Scenario: Customer Support Agent

A long billing conversation is progressively summarized as it grows. By the end the agent refers to the disputed charge as "around forty dollars" instead of the exact $42.17, and the order date drifts to "early March" — causing it to quote the wrong return window.

What practice prevents this loss of precision?

  • A.Maintain a structured "case facts" block (IDs, exact amounts, dates) included in every prompt, independent of how history is summarized.
  • B.Disable summarization entirely and resend the full raw history on every turn.
  • C.Instruct the model to round all amounts to the nearest dollar for consistency.
  • D.Lower the temperature so the numbers stop drifting.

Question 130

Open question ↗

Scenario: Customer Support Agent

Three tools return dates differently: `lookup_order` gives a Unix timestamp, `get_shipment` returns "Mar 5, 2025", and `get_invoice` returns ISO 8601. The agent frequently miscompares dates across them when computing return eligibility.

Which mechanism best normalizes the data before the model sees it?

  • A.A system-prompt instruction telling the model to mentally convert every date to ISO 8601 before comparing.
  • B.A `tool_choice` setting that forces the model to call only one date-returning tool.
  • C.A `PostToolUse` hook that converts every tool result's date fields into one canonical format (ISO 8601).
  • D.Asking each MCP server's vendor to change their output format.