Question 20 A report agent receives partial results from three subtasks, one of which timed out. What should it do? Treat the timeout as a negative finding. Mark that segment incomplete and present only verified conclusions, including the timeout. Report all three as fully completed because two succeeded. Copy the preceding segment’s conclusion into the missing one. Explanation Correct Answer B:Partial aggregation must not imply comprehensive coverage. Incorrect Answers A:No result is not evidence of absence. C:That misstates completeness. D:That creates unsupported content.
Question 21 A one-paragraph FAQ question takes 90 seconds because an orchestrator always runs search, legal review, statistics, and drafting agents. What architectural improvement fits? Run every specialist twice to build consensus. Remove all specialized workers from every future task. Let the coordinator select only task-relevant specialists based on complexity and evidence needs. Move coordinator decisions into isolated workers without shared context. Explanation Correct Answer C:Adaptive delegation avoids paying for a full pipeline on trivial requests. Incorrect Answers A:Double execution increases cost and latency without addressing routing. B:Simple tasks need fewer workers, not a universal ban on specialization. D:Workers then cannot reason about global task requirements or dependencies.
Question 22 A report generator follows a sample format but reproduces a specific fictitious company name in unrelated reports. What should the prompt do? Vary entity values across examples and clearly distinguish format demonstration from factual content. Delete all output structure requirements. Use the same invented company in every example. Ask it to cite the fictitious company as a source. Explanation Correct Answer A:The model should learn the structure without copying sample facts. Incorrect Answers B:The desired format remains valuable. C:Repetition strengthens unwanted literal copying. D:Sample entities are not evidence.
Question 23 Two MCP servers export tools with similar names for different databases. An agent queries the wrong environment. What integration design helps? Place production passwords directly in every tool description. Have both servers return identical unqualified tool names. Assume first-registered always means production. Use unambiguous server-qualified descriptions and allow only the intended environment. Explanation Correct Answer D:Clear names and tool scoping reduce cross-environment routing mistakes. Incorrect Answers A:Exposing secrets is not a routing solution. B:Name collisions make misrouting more likely. C:Registration order is not a safe environment selector.
Question 24 A team wants an AI quality gate without inventing a mandatory human signoff requirement for every low-risk output. What is defensible? Create thousands of review receipts that nobody substantively reads. Claim external expert certification from internal model checks. One substantive model-assisted review with clear severity criteria, targeted fixes and truthful release labeling. Skip all material factual checks to finish faster. Explanation Correct Answer C:A consolidated gate can be proportionate when no external signoff rule applies. Incorrect Answers A:Paperwork alone does not validate content. B:That overstates the assurance achieved. D:Workflow efficiency does not excuse false assertions.
Question 25 A codebase has an undocumented payment subsystem and a proposed refactor that may affect settlements. What approach is safest initially? Explore dependencies and test boundaries in plan mode, then request the needed authorization for edits. Rewrite the settlement service immediately based on filenames. Skip tests because the model can reason about correctness. Run live payment transactions to discover behavior. Explanation Correct Answer A:High-impact uncertain changes benefit from analysis before execution. Incorrect Answers B:Names do not expose all business dependencies. C:Reasoning alone cannot validate system behavior. D:Production side effects are inappropriate for exploratory design.
Question 26 A legal-review model outputs only “high risk” for a vendor contract. What format makes its judgment inspectable? State the clause, exact concern, consequence, confidence, and a suggested redline. Hide uncertainty so the review sounds decisive. Return an unlabeled score without citing a clause. Repeat generic contract best practices regardless of content. Explanation Correct Answer A:Traceable evidence and remedy make the review useful to legal staff. Incorrect Answers B:Uncertainty matters when a clause is ambiguous. C:A number alone does not identify what to change. D:Unconnected advice does not address this agreement.
Question 27 A multi-session research application must allow a user to revisit a final report and inspect the evidence used. What should persistence include? An instruction to remember sources without recording them. Only a final word count. Only transient in-memory references that vanish on restart. Stable report and evidence identifiers linked to the run’s verified findings. Explanation Correct Answer D:A durable artifact map supports reproducibility and safe continuation. Incorrect Answers A:A model promise is not persistent evidence storage. B:Length does not identify sources or results. C:Ephemeral IDs prevent later retrieval and audit.
Question 28 A legal-contract extractor sends all contracts to human reviewers, but most routine clauses are correctly identified. How should the workflow be improved? Use a defined risk rubric to send ambiguous, high-impact or unsupported clauses for review while retaining traceable routine outputs. Ask each reviewer to reread unrelated documents first. Let the model approve legal changes using its own confidence score. Eliminate review for every clause including disputed liability. Explanation Correct Answer A:Risk-based escalation makes reviews focused and auditable. Incorrect Answers B:That adds cost without targeting risk. C:The model does not hold legal approval authority. D:High-impact ambiguity still needs qualified decision-making.
Question 29 A tool returns isError=false with an empty list after a successful search. How does this differ from an isError=true lookup failure? Both cases prove the requested record does not exist. Both cases should be retried forever. A successful empty result means the search ran and found no matches under its query. Both cases authorize creating a new record without consent. Explanation Correct Answer C:The state can support a no-results answer; a tool failure cannot prove absence. Incorrect Answers A:A failed query may never have searched the source. B:A valid no-match can be final unless the search scope changes. D:Neither lookup state grants write permission.
Question 30 A rules file uses paths: ["tests/**"], but generated tests live in packages/service/tests. What should be reviewed? Whether the test files have more than 100 lines. Whether the glob covers nested test directories, then extend it to the intended paths. Whether the agent’s greeting mentions tests. Whether production database tokens are committed. Explanation Correct Answer B:Path patterns must match the actual repository layout. Incorrect Answers A:File length does not determine path matching. C:Greeting content does not load the proper scoped rule. D:Adding secrets is not a glob correction.
