Question 100 A booking MCP tool cannot complete a reservation. Which TWO error fields or properties help the assistant recover accurately? A specific failure category or code distinguishing invalid input from provider outage. A clear flag indicating whether retrying the identical call is appropriate. An empty content array with no error status. A fabricated confirmation number in every error. A request to discard the entire conversation immediately. Explanation Correct Answers A:Different root causes require different next steps. B:Retriable errors can use bounded backoff; permanent ones need changed input or permission. Incorrect Answers C:The agent cannot distinguish failure from an intentionally empty result. D:A false booking reference misleads the traveler. E:The original booking constraints remain needed for recovery.
Question 101 A user asks an assistant for a noncritical draft heading without a preferred tone. What is the most proportionate response? Ask for banking credentials to establish authority. Refuse to write until five independent preferences are collected. Choose a reasonable style and draft it, making revision easy. Treat the missing tone as an instruction to publish publicly. Explanation Correct Answer C:A reversible low-risk formatting choice does not require blocking clarification. Incorrect Answers A:Credentials are irrelevant and sensitive. B:Unnecessary questions delay useful work. D:Drafting and publishing are separate actions.
Question 102 A Python monorepo requires pytest in packages/api, but the frontend must use a different test runner. How should guidance be organized? Delete common conventions and rely on memory alone. Keep shared policies in root guidance and place API-specific instructions near the API subtree. Put Python-only rules globally for all directories. Put both conflicting test instructions into one undifferentiated paragraph. Explanation Correct Answer B:Directory scope limits specialized instructions to relevant files. Incorrect Answers A:Shared rules should remain explicit and reusable. C:Frontend work would inherit irrelevant test commands. D:An unspecific global rule can misroute commands.
Question 103 A prompt designer is selecting examples for a document classification model. Which TWO properties improve generalization? Coverage of representative normal and difficult boundary cases. Examples that ignore the output format required by the API. Hundreds of identical easy examples. Consistent labels that align with the written category definitions. Examples with knowingly incorrect answers as filler. Explanation Correct Answers A:Examples should include the situations where the current decision rule fails. D:Contradictory examples teach incompatible policies. Incorrect Answers B:The model may copy a structure downstream cannot parse. C:Duplication adds context without new decision boundaries. E:Incorrect labels degrade consistency.
Question 104 A worker is expected to use a read-only search tool but receives write permission to the production index. What is the correct configuration? Delete the allowed-tools declaration to allow every tool. Depend solely on a prompt saying never to write. Limit its allowed tools to retrieval and analysis capabilities needed for the assignment. Keep write access so the model can “helpfully” fix data. Explanation Correct Answer C:A worker’s capability set should reflect the narrow delegated task. Incorrect Answers A:Removing constraints does not enforce least privilege. B:Prompt compliance is not a sufficient security boundary. D:The worker was not authorized to mutate the index.
Question 105 A downstream worker receives a tool output with isError set to true but a message body that looks like useful text. What is the correct interpretation? Send the body as an approved external action. Treat the tool response as an error and route it through recovery or escalation, not as verified task data. Ignore isError when the text is grammatical. Replace the result with a fabricated success. Explanation Correct Answer B:Explicit error metadata is part of the result contract. Incorrect Answers A:An error is not authorization. C:Human-readable error text may look like normal content. D:That conceals failure state.
Question 106 A tool backend fails after a response was partially streamed; the agent sees half a JSON record. What should the integration do? Ignore the stream interruption and mark success. Patch missing fields with plausible values automatically. Reject the incomplete payload, surface a structured interruption, and retry only when safe. Always reissue side-effecting calls without reconciliation. Explanation Correct Answer C:Partial data must not be mistaken for a valid complete result. Incorrect Answers A:Incomplete records cannot satisfy the contract. B:Invented data hide source incompleteness. D:Unsafe retries can duplicate external actions.
Question 107 A company wants consistent concise code-review comments, but the few-shot prompt uses lengthy essay-style outputs. What should change? Only increase maximum output tokens. Use examples matching the required finding schema, brevity, and evidence detail. Remove all evidence fields to force short answers. Add more long essays to emphasize thoroughness. Explanation Correct Answer B:Few-shot output form should resemble the intended production response. Incorrect Answers A:Capacity does not fix the exemplar style. C:Brevity should not eliminate actionable detail. D:Those train the wrong format.
Question 108 An automated job resumes an agent session but the underlying repository has changed since its last tool result. What must it do before editing? Discard the session and all approved decisions automatically. Treat every cached file view as permanently current. Overwrite conflicting files without checking differences. Revalidate external workspace state and dependencies before applying stale planned edits. Explanation Correct Answer D:Session history preserves past observations, not a guarantee that mutable files are unchanged. Incorrect Answers A:State changes require targeted revalidation rather than mandatory restart. B:Another developer may have modified those files. C:Blind overwrites can destroy intervening work.
Question 109 A subordinate reviewer tells a writer that a citation is verified, but the URL points to an unrelated article. What should the writer do? Open the source and verify it supports the exact claim before reusing the reviewer’s assertion. Prefer whichever source title sounds more authoritative. Copy the citation because the reviewer supplied it. Remove all source names and still label the claim proven. Explanation Correct Answer A:Delegated confidence is not a substitute for source–claim alignment. Incorrect Answers B:Authority is insufficient without relevance. C:Incorrect citations propagate unsupported claims. D:That hides the problem.
Question 110 A weather MCP tool times out because a backend is temporarily unavailable. Which response best guides the agent? A structured transient error with retry guidance and bounded backoff information. An empty success payload. A made-up forecast to preserve conversational flow. A permanent invalid-argument error despite valid input. Explanation Correct Answer A:A temporary outage may be recoverable without changing the user’s request. Incorrect Answers B:The agent could mistake lack of data for a clear forecast. C:Fabricated data are unacceptable substitutes for unavailable observations. D:Misclassification would discourage useful bounded retries.
