Question 20 An agent sends incident notifications to several departments. Which integration requirement should be documented? Assume message delivery is always exactly once Never log delivery results Explicit recipients, event triggers, deduplication and retry policies Send messages on every user utterance Explanation Correct Answer C:Notification contract details prevent duplicate or missed incident messages. Incorrect Answers A:Distributed systems require retry and duplicate handling. B:Without logs failed notifications cannot be traced. D:Unfiltered triggers could create alert storms.
Question 21 A bank wants a model to recommend loan decisions but humans remain legally accountable. Which architecture reflects this? Allow agent outputs to change loan status automatically Suppress explanatory evidence to avoid bias discussions Use model certainty as a substitute for policy compliance Present evidence and recommendations to authorized reviewers before final decision Explanation Correct Answer D:Human review provides accountability for a high-impact regulated outcome. Incorrect Answers A:Automatic adjudication bypasses the stated approval boundary. B:Missing rationale prevents meaningful review and challenge. C:Confidence does not establish regulatory conformity.
Question 22 An internal IT agent should be usable only by the cybersecurity team, not the whole tenant. Which step belongs in deployment planning? Provide a shared administrator login in documentation Publish to an anonymous website and rely on user discretion Configure authentication and restrict agent sharing to the intended organizational audience Add a hidden system prompt saying 'security only' Explanation Correct Answer C:A private channel alone may not restrict discoverability or use. Incorrect Answers A:Shared privileged credentials defeat access restrictions. B:Anyone with the link could interact without team membership checks. D:Prompt text is not an audience access control.
Question 23 An agency needs reusable request validation across a billing agent and helpdesk agent. Which component boundary is appropriate? Disable input validation and rely on model judgment Copy validation examples into each welcome prompt Create a versioned shared validation tool with defined inputs and outputs Store validation logic only in unstructured past chat Explanation Correct Answer C:A reusable contract can serve multiple agents and be tested centrally. Incorrect Answers A:Free-form judgment may allow malformed external requests. B:Examples do not perform data validation consistently. D:Conversation history is not a deployable service contract.
Question 24 An agent makes calls to an inventory system that sometimes times out. What should the design include? Pretend an order was accepted when the API fails Repeat write requests endlessly without idempotency Timeout boundaries, safe retries and a user-facing failure or escalation path Delete the order attempt from all telemetry Explanation Correct Answer C:External outages require deterministic recovery and honest status. Incorrect Answers A:Fabricating success risks missing or duplicated orders. B:Unbounded retries can create duplicate state changes. D:Lost evidence prevents incident diagnosis.
Question 25 A child-facing learning agent might be asked for financial account changes. What is the first planning principle? Treat any typed adult name as parental permission Keep age-appropriate educational functions separate from protected financial transactions Give all learners unrestricted account-management tools Hide financial tool calls from the parent Explanation Correct Answer B:Sensitive account actions require additional authorization and eligibility safeguards. Incorrect Answers A:Text alone cannot verify the account holder or guardian. C:This expands a teaching agent into high-risk financial operations. D:Unobserved changes create accountability and consent problems.
Question 26 An employee agent mixes knowledge from multiple subsidiaries. Why document source ownership and classification? To let the model ignore department permissions To allow public sharing when queries are popular To remove all data expiration rules To ensure each source follows its own permission and data-handling policy Explanation Correct Answer D:Shared indexing must not dissolve subsidiary boundaries. Incorrect Answers A:Central retrieval does not override source restrictions. B:Popularity does not change confidentiality classification. C:Classified content still needs lifecycle governance.
Question 27 An agent recommends chemical handling procedures and may be uncertain. Which responsible AI plan is appropriate? Generate unapproved operating procedures on demand Use verified safety documents and hand off unsupported safety-critical requests to qualified personnel Hide knowledge source provenance from operators Optimize for fast answers at the expense of accuracy Explanation Correct Answer B:Unsafe guesses require escalation instead of invented instructions. Incorrect Answers A:Unverified steps can create physical hazards. C:Operators need traceable authoritative instructions. D:Latency alone is not a safety control.
Question 28 A multi-department agent will call both a read-only directory tool and a payment-approval tool. Which permissions strategy is safer? Disable identity checks for read-only features and payment alike Give all authenticated users payment administrator tokens Use the same broad service account for every tool Provide narrowly scoped tool permissions and separate high-risk approval rights Explanation Correct Answer D:A payment tool needs stronger rights than a directory search. Incorrect Answers A:Payment access still requires user-specific authorization. B:Authentication alone does not confer purchase authority. C:Shared privileges amplify any tool misuse.
Question 29 A legal team wants to reuse the same reviewed disclaimer in agents on multiple channels. Which architecture helps control its wording? Allow each agent owner to adjust the legal disclaimer in its local prompt Maintain a governed approved response component and version its use across agents Have every agent paraphrase the disclaimer freely Copy an outdated disclaimer from chat examples Explanation Correct Answer B:Central review reduces inconsistent legal wording. Incorrect Answers A:Local editing would diverge from the single legally reviewed wording across channels. C:Paraphrasing can remove mandatory language. D:Historical samples are not authoritative approved content.
Question 30 A factory agent calls operational technology APIs that may affect equipment. Which risk assessment is essential? Expose equipment writes to public guest conversations Plan network isolation, explicit operator approval and safe failure behavior for physical actions Let the agent bypass plant safety interlocks Rely only on polite user instructions Explanation Correct Answer B:Production equipment changes have safety and availability consequences. Incorrect Answers A:Anonymous users should not control industrial systems. C:Software convenience does not justify bypassing physical safety. D:Prompts do not enforce safe operational states.