Question 590
An agent calls multiple tools, and reviewers cannot link its final answer to the tools used. What evaluation evidence should be collected?
Explanation
Correct Answer
Tool-chain evidence supports checking whether the agent relied on authorized, correct intermediate operations.
Incorrect Answers
Add adversarial and boundary cases for disallowed tool use and missing approvals. This is useful for the distinct issue: An agent can initiate customer refunds, but the test dataset contains no unauthorized-tool scenarios. The relevant technical rationale or evidence for that issue is: High-impact tool behavior requires negative tests proving the agent respects authorization and confirmation rules. Here the reported issue is: An agent calls multiple tools, and reviewers cannot link its final answer to the tools used. The relevant response is: Capture tool invocation traces and outputs alongside the final response and user query. The first result cannot verify this different correction: Tool-chain evidence supports checking whether the agent relied on authorized, correct intermediate operations.
Minimize and deidentify sensitive records while preserving the evaluation's necessary characteristics. It addresses another case: A test collection accidentally contains customer email addresses and account numbers that need not be present. For that case, the distinguishing justification or acceptance check is: Test quality does not require uncontrolled personal identifiers; privacy safeguards should apply to evaluation corpora as well. The present case instead states: An agent calls multiple tools, and reviewers cannot link its final answer to the tools used. It calls for Capture tool invocation traces and outputs alongside the final response and user query. Showing success for the other case would not establish: Tool-chain evidence supports checking whether the agent relied on authorized, correct intermediate operations.
The alternative, Build a representative evaluation dataset including ordinary and ambiguous high-risk user intents, belongs to a different technical condition: A RAG assistant is evaluated on only easy factual queries even though it also handles ambiguous refund requests. There the source-specific mechanism or proof is: Test cases must reflect the variety of real requests; easy-only evaluation masks failures on important edge cases. The condition asked about here is: An agent calls multiple tools, and reviewers cannot link its final answer to the tools used. A suitable correction is Capture tool invocation traces and outputs alongside the final response and user query; the targeted justification or result is: Tool-chain evidence supports checking whether the agent relied on authorized, correct intermediate operations.