Question 1 A tool-led workflow returned a successful payment ID, but the conversation is too long to retain every message. What is the right durable state? Persist the exact verified payment ID and outcome with its source receipt reference. Repeat the payment after compaction because the prior conversation is gone. Store only the number of tool calls made. Regenerate a payment ID from the customer name later. Explanation Correct Answer A:A compact, verifiable artifact preserves the completion fact without replaying the whole chat. Incorrect Answers B:The operation may already have succeeded. C:Call counts do not identify the payment. D:Identifiers cannot be safely guessed.
Question 2 An agent tries to search a catalog by using the order-placement tool. What description should the placement tool include? State that it creates a real order and requires a previously confirmed purchase decision. Remove permission requirements from the documented interface. Claim order placement is a read-only operation. Tell the agent to call it for every product-related question. Explanation Correct Answer A:Explicit side effects and prerequisites differentiate checkout from discovery. Incorrect Answers B:Documentation must not erase the approval boundary. C:That misrepresents a financial side effect. D:Broad routing would trigger orders for informational queries.
Question 3 A monorepo needs frontend-specific lint conventions and infrastructure-specific security requirements. Which TWO rule-design choices are appropriate? Copy production credentials into the rules to reduce setup friction. Write distinct rules with explicit paths globs for frontend and infrastructure subtrees. Test representative matching and nonmatching file paths for each rule. Place every rule in one unconditional global block. Use broad patterns matching all directories regardless of purpose. Explanation Correct Answers B:Scope controls keep unrelated instructions from colliding. C:Pattern verification catches accidental overreach before automated edits. Incorrect Answers A:Rules files should not embed secrets. D:This loads irrelevant guidance in every task. E:Overinclusive globs undermine conditional loading.
Question 4 A parent agent receives results in a different order from the sequence it launched subagents. Which design avoids ordering assumptions? Drop any result that arrives out of sequence. Sort by response length and assume longer results belong to complex jobs. Match responses to outstanding jobs by explicit task identifiers, not arrival order. Assume the first launched job always returns first. Explanation Correct Answer C:Parallel completions are nondeterministic; identity mapping makes aggregation robust. Incorrect Answers A:Valid work would be lost merely because it completed faster or slower. B:Text size does not identify the originating task. D:Different runtimes and external calls invalidate that assumption.
Question 5 A quality-review agent repeatedly reports stylistic nits as critical bugs. What should its second-pass rubric enforce? Treat all suggestions as equally urgent. Forbid the reviewer from reading the source. Require reproducible defect evidence, consequence and severity rationale for each blocker. Require at least twenty critical findings per change. Explanation Correct Answer C:Evidence criteria distinguish true release risk from subjective preference. Incorrect Answers A:Severity should reflect risk and impact. B:The review would become ungrounded. D:Finding quotas inflate false positives.
Question 6 A compliance memo cites a policy that was superseded by a new revision. What is the right provenance response? Identify the effective policy version and note any relevant change from the superseded edition. Prefer the older policy because its PDF is shorter. Treat a superseded document as automatically current. Cite both without distinguishing effective dates. Explanation Correct Answer A:Version provenance determines which rule actually applies. Incorrect Answers B:Length has no authority significance. C:Availability is not applicability. D:That can create conflicting duties.
Question 7 A model repeatedly modifies the wrong function because two symbols share the same name. What should it do before the next Edit? Delete both functions indiscriminately. Repeat the same global replacement with higher confidence. Inspect call sites and file context, then target the exact implementation. Assume the first Grep hit is always authoritative. Explanation Correct Answer C:Precise context resolves ambiguity before another mutation. Incorrect Answers A:Destructive removal ignores the intended behavior. B:Unchanged targeting will likely repeat the error. D:Result order does not establish the correct function.
Question 8 An agent tries to call delete_record on an account outside its permitted tenant. What is the best hook placement? Check only after the deletion has succeeded. Run a chat summary hook after every 20 turns. Intercept the proposed tool call before execution and reject cross-tenant identifiers. Add a logging-only hook that never blocks calls. Explanation Correct Answer C:Pre-execution authorization blocks the unsafe mutation before it occurs. Incorrect Answers A:A post-call inspection cannot guarantee recovery of deleted records. B:Summarization is unrelated to tenant authorization. D:Audit logs cannot enforce a deny requirement by themselves.
Question 9 An MCP write tool rejects an unauthorized account update. Should the model retry with the same arguments? Yes, retry rapidly until authorization succeeds. Replace the account identifier with a random nearby ID. No; return an explicit nonretryable authorization error and request appropriate access. Mark the update complete because the service responded. Explanation Correct Answer C:Repeating a denied mutation does not create permission. Incorrect Answers A:Repeated denied calls can trigger abuse controls and still violate policy. B:Guessing another target could cross an access boundary. D:A response carrying denial is not an executed update.
Question 10 A model extracts dates consistently from English invoices but fails on European day-month formats. What should examples add? Only more US-format dates. A direction to ignore ambiguous dates silently. Locale-diverse dates with normalized ISO outputs and explicit ambiguity handling. An arbitrary default month for every number. Explanation Correct Answer C:Examples should cover the format variation causing errors. Incorrect Answers A:Those do not teach day-month parsing. B:Silent omission hides source uncertainty. D:Guessed conversions can be wrong.
