Question 200 An agent’s tool list contains two tools with the same side effect but different billing permissions. How should it avoid privilege confusion? Rotate the display order daily and hope the model learns. Describe both as unrestricted billing updates. Always select the more privileged tool. Name and document the authorization boundary explicitly, and expose only the appropriate tier to each agent. Explanation Correct Answer D:Tool scope and descriptions must distinguish restricted from privileged operations. Incorrect Answers A:Tool ordering is not a permission mechanism. B:Misleading descriptions erase the security difference. C:Privilege should be minimized to what the operation needs.
Question 201 After a first-pass automated review, a team wants robust refinement without needless churn. Which TWO actions are most useful? Restrict the next edit to code involved in the failing behavior and rerun relevant checks. Turn concrete failing examples into reproducible tests or assertions. Remove tests that disagree with the model’s initial answer. Treat the first passing command as proof of all requirements. Rewrite every file to a new style before debugging. Explanation Correct Answers A:Focused changes reduce regression risk and provide clear validation. B:Test fixtures make the defect observable across future iterations. Incorrect Answers C:Suppressing feedback prevents real correction. D:One check may not cover the specific regression. E:Large unrelated diffs obscure the failing behavior.
Question 202 An assistant is preparing a high-impact external account change under an ambiguous request. Which TWO checks are appropriate? Assume the most privileged target when several matches exist. Verify that the action is permitted under the available role and policy before execution. Use credentials from a different tenant to bypass a permission error. Confirm the exact target account and requested state against the user’s authorization. Send a success report before executing the change. Explanation Correct Answers B:User intent and system authorization are distinct constraints. D:Correct target identification is critical for irreversible operations. Incorrect Answers A:That can change the wrong account. C:That violates account separation. E:Status must reflect actual tool outcomes.
Question 203 An agent proposes to send an email containing confidential attachments. The policy requires a classification and recipient check before dispatch. Where is the safest enforcement point? Only in a prompt reminding the model that confidentiality is important. After mail delivery, by asking recipients to delete it. In the send-email execution path, validating labels and authorized recipients before delivery. In an unrelated backup process that runs overnight. Explanation Correct Answer C:The dispatch boundary can block disclosure regardless of the model’s proposed wording. Incorrect Answers A:Prompt guidance cannot guarantee every destination and attachment is checked. B:An external disclosure cannot reliably be recalled. D:Backup handling does not authorize or prevent the outgoing email.
Question 204 A batch pipeline submits 10,000 records twice after a coordinator restart. What prevents duplicate downstream entries? Assume retries are never needed in asynchronous systems. Depend on exact response arrival order. Persist request identity and make database writes idempotent against stable business keys. Generate new keys for every retry. Explanation Correct Answer C:Retries and restarts should not duplicate successful operations. Incorrect Answers A:Recoverable failures are normal. B:Order is not a uniqueness guarantee. D:New keys remove duplicate detection.
Question 205 An agent must choose among four equally named tools but only one is relevant. What is the best first repair? Improve descriptions and scope tool availability to the actual task rather than globally exposing all four. Alphabetize tool registrations and call the first. Remove all arguments from every tool. Ask the model to choose randomly for fairness. Explanation Correct Answer A:Better semantics and smaller candidate sets reduce misrouting. Incorrect Answers B:Tool order is not evidence of relevance. C:Loss of input schemas worsens integration reliability. D:Random selection is unsuitable for side-effectful APIs.
Question 206 A path-specific security review rule is loaded only after a sensitive file has been written. What should be fixed? Move the rule to a post-release announcement. Disable the rule because it causes extra work. Ensure the matching rule is present before edits and gate sensitive writes if necessary. Trust a later summary to undo any leaked secrets. Explanation Correct Answer C:Instructions and safety constraints must influence the action before it occurs. Incorrect Answers A:Late communication cannot prevent the original unsafe edit. B:Removing a relevant security rule increases exposure. D:A leak may persist outside the current run.
Question 207 A client receives stop_reason end_turn following synthesis and no tool requests remain. What should its agent loop do? Continue polling the model until it uses a tool. Invent an empty tool_result to acknowledge completion. Restart with an empty system prompt for a final rewrite. Return the assistant’s completed response to the caller. Explanation Correct Answer D:end_turn indicates the assistant has finished this turn; no further tool exchange is needed. Incorrect Answers A:Additional turns after completion create unnecessary work and may change the answer. B:No tool request exists to pair with a tool_result. C:Discarding context risks changing the verified synthesis.
Question 208 The worker processes a batch, but an upstream source file changes before results are committed. How should consistency be maintained? Assume document names uniquely identify immutable content. Record input versions or hashes and reject results that no longer match the authoritative source version. Remove all version fields to simplify the interface. Apply all old results regardless of changes. Explanation Correct Answer B:Version binding prevents stale output from overwriting newer information. Incorrect Answers A:A filename may refer to changed bytes. C:The consumer loses staleness checks. D:That can replace current data with stale classifications.
Question 209 An agent must hand off a partly completed database cleanup. Which TWO facts are essential to preserve? The assistant’s preferred font for status reports. The exact rows already changed and the transaction or batch identifier. A random sample of unrelated records from another database. An invented success status to simplify the handoff. The remaining approved scope and any unresolved validation failures. Explanation Correct Answers B:Completed work must be identifiable to avoid double mutation. E:Continuation must know what is authorized and what is blocked. Incorrect Answers A:Presentation preference does not restore database state. C:It does not describe this job’s progress. D:False closure risks data loss.
Question 210 A tool calls an external service that is throttling. What should a structured response preserve? A fake successful record to hide capacity problems. Only a generic “bad” string. An instruction to remove tool request limits globally. A recognizable rate-limit category and a safe retry-after interval when available. Explanation Correct Answer D:The caller can delay and cap future calls rather than hammering the backend. Incorrect Answers A:False output pollutes downstream results. B:The agent cannot distinguish waiting from correcting its input. C:Bypassing rate limits is not a valid recovery.
