Question 390 A model requests a database update and a read in the same tool batch, but the read depends on the update. What should the host ensure? Respect the dependency and run the read only after the authorized update completes. Read first and report the stale value as current. Execute both concurrently because multiple tool uses always imply independence. Skip the read and assert the update succeeded. Explanation Correct Answer A:Parallelizing dependent operations can produce stale or inconsistent results. Incorrect Answers B:The reading may not reflect the intended update. C:A common response batch does not guarantee there are no data dependencies. D:A submitted update still needs the appropriate outcome check.
Question 391 A workflow needs to deny shell commands containing unauthorized path traversal. Which hook must be positioned before the shell operation? A post-tool report formatter that renames the path. A pre-tool validation hook that canonicalizes and checks the destination path. A Stop hook that fires after the user sees the answer. A conversation summarizer that truncates path names. Explanation Correct Answer B:Validation after canonicalization can block unauthorized targets before execution. Incorrect Answers A:Renaming output text cannot undo a prohibited filesystem operation. C:The unsafe command would already have run. D:Shortening names is not permission enforcement.
Question 392 A PostToolUse hook receives malformed JSON from an external CRM. What should it do? Pass an empty success object and allow a refund. Return a structured, explicit parse error rather than fabricating fields. Guess the missing values from similar customer records. Disable the hook whenever input is malformed. Explanation Correct Answer B:A hook must not invent customer state when normalization fails. Incorrect Answers A:Downstream logic may mistake an empty object for verified information. C:This can cross-contaminate accounts and create false facts. D:Skipping validation removes the control exactly when it is needed.
Question 393 A pipeline requires document classification before routing classified files to different analyzers. What task graph is appropriate? Run the classifier first, then dispatch each file to an analyzer selected from the classification. Randomly select an analyzer to reduce coordinator overhead. Run all analyzers before classifying any file. Merge classifier and report generator while skipping analysis. Explanation Correct Answer A:Downstream routing depends on an upstream label; classification is a prerequisite. Incorrect Answers B:Random routing does not respect document types or capability boundaries. C:Unnecessary or inappropriate analyzers may process documents before their type is known. D:The system still needs the specialized analysis step.
Question 394 A deployment agent must change a firewall and then prove the change did not cut off production. Which TWO gates are essential? Run a post-change connectivity check and invoke rollback on failure. Require approved change scope before writing firewall policy. Disable management logging until the change is over. Treat a successful API response as proof all applications work. Allow the agent to expand the change to any network it discovers. Explanation Correct Answers A:Verification and backout prevent a silent service outage. B:The agent should not alter network rules outside authorized scope. Incorrect Answers C:Lost logs make failures harder to diagnose and audit. D:Configuration acceptance does not guarantee end-to-end service health. E:Discovery is not permission for scope expansion.
Question 395 A document agent says “done” in a text block while stop_reason is tool_use and its search call has not run. Which signal should the orchestration loop trust? Stop whenever one tool call succeeds. Stop when the output reaches 500 tokens. Stop as soon as the response contains the word done. Use stop_reason and execute pending tool_use blocks before considering completion. Explanation Correct Answer D:The protocol stop reason identifies a tool handoff; words inside generated text are not a completion protocol. Incorrect Answers A:A successful tool can be followed by more needed actions or synthesis. B:Token length measures size, not whether the task has finished. C:Natural-language wording is unreliable and would skip required tools.
Question 396 A coordinator passes customer PII to every specialist even though only one needs it. What data-minimization strategy should it use? Include full customer histories in all prompts as a precaution. Provide each worker only the fields necessary for its authorized task. Mask only the final public report after workers have seen the PII. Ask specialists to remember data but not mention it. Explanation Correct Answer B:Reducing sensitive context lowers exposure while preserving task performance. Incorrect Answers A:Unneeded data increases privacy risk and token consumption. C:Final output redaction does not undo upstream overexposure. D:Instructions cannot remove unnecessary access to the data.
Question 397 A legal specialist has access to confidential contracts, but a web-research specialist only needs public information. How should the orchestrator handle capabilities? Let the web researcher borrow the legal worker’s token by default. Place raw secrets in every worker’s prompt for convenience. Expose confidential tools only to the authorized legal specialist. Give every worker all tools to prevent capability errors. Explanation Correct Answer C:Least-privilege delegation reduces the chance of inappropriate data exposure. Incorrect Answers A:Credential sharing bypasses the intended security boundary. B:Shared prompts are a poor secret distribution mechanism. D:Broad access creates unnecessary privilege and data-leak pathways.
Question 398 Two analysts must independently assess the same safety report to detect missed hazards. What decomposition preserves independence? Give reviewer B only reviewer A’s conclusion. Accept any one reviewer’s lowest-risk rating automatically. Run independent reviews from the same source, then adjudicate disagreements with evidence. Merge all reviewers’ scratchpads during the initial analysis. Explanation Correct Answer C:Separate first passes reduce anchoring while a controlled merge resolves conflicts. Incorrect Answers A:A second review would no longer be independent. B:Unresolved discrepancies require examination rather than arbitrary selection. D:Shared reasoning can propagate one worker’s omissions.
Question 399 A workflow plans to resume an old session by embedding a user-provided string in an SDK resume parameter. What should it check? Assume anyone who knows an ID owns the session. Log the entire session including secrets to public telemetry first. Disable tenant isolation to make resumes easier. Validate the session ID against the user’s authorized stored sessions. Explanation Correct Answer D:A supplied identifier must not grant access to another user’s conversation state. Incorrect Answers A:Identifiers can leak and should not be the only authorization control. B:Overbroad logs can expose sensitive state. C:That would allow cross-tenant state access.
Question 400 A data specialist returns a JSON table but the parent expects a quoted source passage and page number. What should the integration contract specify? A typed handoff schema with required evidence fields and validation on receipt. A loose natural-language promise to return “something useful.” An arbitrary limit of ten words per response. Automatic acceptance of malformed handoff messages. Explanation Correct Answer A:Producer and consumer must agree on claim text, source, location, and completeness. Incorrect Answers B:Unspecified output structure yields recurring integration failures. C:A word cap cannot guarantee the necessary citation fields. D:Accepting invalid records contaminates later synthesis.
