Question 70 An agent drafts a routine internal summary with source links and no external side effects. When is human intervention most proportionate? Require personal approval for every punctuation mark. Publish it externally without checking authorization. Treat source links as guaranteed factual correctness. When a material unsupported claim or policy-triggering uncertainty is identified, not for every harmless sentence. Explanation Correct Answer D:Risk-based review keeps the workflow efficient while catching significant errors. Incorrect Answers A:Microapprovals create friction without stronger assurance. B:A draft does not authorize public release. C:Links must support the specific assertions.
Question 71 A user asks “What would you change?” about authentication architecture, and the agent opens a file-writing tool. What is the error? It chose a working directory with a short name. It did not call the deployment tool first. It crossed from analysis into execution without a request to modify the code. It used too few words in the final response. Explanation Correct Answer C:A design recommendation can be satisfied with read-only exploration and an explicit proposal. Incorrect Answers A:Path length is unrelated to user intent. B:Deployment would be an even greater unauthorized action. D:Length does not determine action authorization.
Question 72 A design agent must compare two prompt strategies from the same initial data without letting one strategy influence the other. Which setup is appropriate? Change success criteria after seeing each strategy’s result. Run the same branch twice without identifying which configuration was used. Use one session where strategy B reads strategy A’s hidden draft. Fork the initial state into separate evaluation branches and compare against shared criteria. Explanation Correct Answer D:Independent branches reduce cross-contamination while retaining a common baseline. Incorrect Answers A:Moving evaluation criteria biases comparison. B:Unlabeled runs cannot support a meaningful comparison. C:This does not preserve independent evaluation.
Question 73 A model-based critic identifies a suspected dependency vulnerability with no package version or advisory. What should the main agent do? Delete the critic’s finding without review. Report the suspicion as an exploit already observed. Upgrade unrelated dependencies blindly. Verify the package and affected version against a relevant advisory before treating the claim as confirmed. Explanation Correct Answer D:Independent factual evidence is needed to avoid speculative vulnerability findings. Incorrect Answers A:Potential material risk still merits verification. B:The critic did not provide exploit evidence. C:That introduces risk without resolving a proven issue.
Question 74 A tool named get_account sometimes returns an account and sometimes silently creates one. What is the main design flaw? The tool is invoked by an LLM rather than a web browser. A read-looking operation has a hidden write side effect; split or clearly rename the operations. The account identifier might be an integer. The response takes too few tokens. Explanation Correct Answer B:Tool names and descriptions should communicate mutations so routing and approval are reliable. Incorrect Answers A:Client type is not the reason an operation must declare its effects. C:Identifier type does not explain the hidden side effect. D:Response size does not make implicit account creation safe.
Question 75 A team plans to use Claude Code for a high-risk database migration. Which TWO steps should precede direct implementation? Run migration scripts on production immediately to see what happens. Inspect dependencies and data invariants in a nonmutating planning phase. Assume that a detailed prompt guarantees a safe migration. Agree on a test and rollback strategy before executing writes. Delete backups to simplify the workspace. Explanation Correct Answers B:The plan must establish what could break and what must be preserved. D:Validation and backout reduce operational risk from a failed migration. Incorrect Answers A:Exploration should not rely on uncontrolled live side effects. C:Prompts do not replace actual controls and tests. E:Backups may be essential for recovery.
Question 76 A developer requests an explanation of how messages flow through a complex system, not a code change. Which output is appropriate? A raw dump of every file in the repository. A sweeping refactor because the architecture seems unusual. A speculative architecture diagram that omits all source paths. A source-linked path from entrypoint through relevant handlers and queues, noting unverified branches. Explanation Correct Answer D:A traceable architecture map answers the research request without editing. Incorrect Answers A:Volume is not an intelligible explanation. B:The user requested explanation, not modification. C:It cannot be validated against the repo.
Question 77 A coding assistant repeatedly generates commands without observing their effects, and its plan becomes detached from the repository. What feedback design is needed? Ask the model to hallucinate the expected command outputs. Set a very high temperature for additional creativity. Return actual execution outcomes to the conversation before selecting the next action. Disable all status and error reporting from tools. Explanation Correct Answer C:An autonomous loop depends on observing results, not inventing state after commands. Incorrect Answers A:Simulated outcomes conceal failures and make later actions unsound. B:Randomness cannot substitute for observing repository state. D:Without results the model cannot adapt its plan to reality.
Question 78 Some batch requests are repeatedly rate limited. What scheduling change is most suitable? Classify rate-limit responses as successful predictions. Increase simultaneous retries until one succeeds. Drop rate-limited records without tracking them. Use documented capacity limits, controlled submission rates and bounded backoff on retryable failures. Explanation Correct Answer D:Scheduling must account for service capacity rather than generating a retry storm. Incorrect Answers A:Transport errors are not model outputs. B:Excess pressure can worsen throttling. C:That creates silent gaps.
Question 79 A model can access search_orders and refund_order. The user asks “Where is my package?” What should its tool distribution and choice encourage? Use both tools on every order request. Read-only order status lookup, with refund mutation unavailable absent an authorized refund request. Answer from a remembered order placed last year. Refund first because it may improve user satisfaction. Explanation Correct Answer B:The requested goal is informational, not a financial transaction. Incorrect Answers A:Mandatory mutations would be inappropriate for routine tracking. C:Current shipping status requires fresh authorized data. D:A shipping question is not permission to reverse a charge.
Question 80 A market-research assistant cites a webpage for a numerical claim, but the cited page never mentions that number. What is the correct treatment? Keep the number because a URL is present. Remove or qualify the claim until an actually supporting source is found. Invent a quotation that would justify the value. Replace the citation with an unrelated reputable homepage. Explanation Correct Answer B:A citation is valid only when it substantiates the specific statement. Incorrect Answers A:A link is not automatically evidence. C:Fabricated quotes worsen the error. D:Reputation alone does not support the figure.
