Question 130 A deployment agent can run tests and propose infrastructure edits, but must not secretly publish a release. Which TWO tool controls are appropriate? Restrict available write tools to the authorized workspace and environment. Provide every production deployment credential in the prompt. Treat any passing unit test as unconditional deploy consent. Hide the tool actions from operators to avoid interruptions. Require explicit release approval before exposing or executing the deploy capability. Explanation Correct Answers A:Tool scoping limits the side effects the agent can attempt. E:A separate authorization gate protects production publication. Incorrect Answers B:Prompt access to secrets bypasses controlled tool boundaries. C:Passing tests are not a release authorization. D:Lack of visibility undermines auditability and oversight.
Question 131 An enterprise wants to compare two model configurations over the same sampled document population. What experimental batching practice is defensible? Use different, unrecorded document samples for each model. Ignore failures from one configuration. Run both configurations against an identical versioned input set and compare a fixed rubric with source-linked results. Judge only by which response is longer. Explanation Correct Answer C:Matched inputs isolate configuration effects and support reproducible evaluation. Incorrect Answers A:Sample differences confound quality comparisons. B:Excluding failures biases results. D:Length is not correctness.
Question 132 An organization wants to record that security-sensitive source edits must be reviewed, while an individual prefers short replies. What is a suitable split? Put the mandatory review rule only in one developer’s home directory. Put the personal style preference into a mandatory organization policy. Store neither rule and assume agents remember them across sessions. Shared project guidance for the review requirement; user-level guidance for personal answer style. Explanation Correct Answer D:Different authority and audience call for separate memory scopes. Incorrect Answers A:Other contributors would miss the project control. B:A subjective reply style need not bind the whole team. C:Ephemeral model context does not guarantee policy persistence.
Question 133 A coordinator has a predictable daily ETL workflow, but engineers want it to plan all steps anew on every execution. What should the architect prefer? Let the agent rewrite database schemas automatically each day. Use open-ended planning for every fixed transformation. Encode known sequencing and validations, reserving agent judgment for genuine exceptions. Remove data validation to accelerate daily completion. Explanation Correct Answer C:Deterministic routines are cheaper, easier to test, and less prone to drift. Incorrect Answers A:Schema mutations are outside normal ETL planning and create risk. B:Replanning a stable pipeline adds uncertainty without benefit. D:Reliability controls are still required even in predictable tasks.
Question 134 A developer requests creation of a brand-new markdown runbook that does not exist. Which tool is appropriate once path and content are decided? Write to create the new file with the intended content. Edit a different existing runbook arbitrarily. Read a nonexistent file and call it delivered. Grep for the file name without writing anything. Explanation Correct Answer A:A creation operation needs an explicit file write. Incorrect Answers B:That changes the wrong artifact. C:Reading cannot create the artifact. D:Search does not satisfy file creation.
Question 135 A large batch includes a few invalid requests but most completed successfully. What is the safest follow-up? Reconcile each request status and resubmit only failed eligible IDs after correcting their errors. Treat one failure as proof all returned successes are false. Repeat every completed request without checking prior outcomes. Discard the failed items silently. Explanation Correct Answer A:Per-item accounting avoids duplicate work and lost outputs. Incorrect Answers B:Independent work items can complete separately. C:That wastes cost and risks duplicate downstream effects. D:A complete job must account for every input.
Question 136 A custom review command must run in an isolated subagent context rather than fill the main conversation with tool output. Which configuration concept fits? Use the supported skill or command fork-context configuration for that workflow. Disable all subagent tools so no review can run. Add raw tool results to the main conversation indefinitely. Commit confidential session state into a public README. Explanation Correct Answer A:An isolated execution context can return a bounded result to the main session. Incorrect Answers B:The workflow still needs its authorized tools. C:That defeats the context isolation goal. D:Unnecessary disclosure does not create safe isolation.
Question 137 A public dataset was last updated four years ago, but an agent writes a present-tense claim about today’s market share. What should change? Hide the dataset year in a footnote. Keep present tense because the old dataset is public. Date the historical observation and avoid claiming current market share without fresh evidence. Extrapolate an exact new share without a model or data. Explanation Correct Answer C:Time-sensitive facts must be aligned with source freshness. Incorrect Answers A:That obscures the central limitation. B:Public availability does not imply current validity. D:An unsupported current number is misleading.
Question 138 A team must review both the security of five microservices and the data flows between them. Which decomposition is best? Merge the five services into a single context-free checklist. Inspect only network traffic and ignore authentication logic. Inspect only service-local code, assuming interfaces are secure. Inspect each service’s local controls and separately examine shared interfaces and trust boundaries. Explanation Correct Answer D:Both local components and cross-service data flows can introduce distinct risks. Incorrect Answers A:A generic checklist loses per-service and boundary specifics. B:Transport observations cannot replace service-level authorization review. C:Many security failures appear at boundaries rather than inside isolated services.
Question 139 An MCP tool wraps an error message as an ordinary success text block. The model later cites the error as a search finding. What contract defect needs repair? Represent failure in the tool result’s error field rather than only as untyped text. Always return a placeholder report with invented facts. Ask the final writer to ignore any sentence containing “error.” Remove the error message but leave success status. Explanation Correct Answer A:Explicit error signaling prevents operational messages from masquerading as evidence. Incorrect Answers B:Fabrication does not fix a failure-status contract. C:Keyword filtering is not a reliable interface contract. D:Hiding the failure would make the misinterpretation worse.
Question 140 A model returns dates as strings that conform to the schema but schedule an event before the contract signing date. What must be added? A temporal business-rule validation that enforces the permitted date ordering. An assumption that schema-valid dates must be legal. A check that every date contains two hyphens. A rule that all dates are replaced with today. Explanation Correct Answer A:Cross-field invariants go beyond individual field types. Incorrect Answers B:Schema validation does not enforce contract semantics. C:String shape cannot establish chronology. D:That discards the source information.
