Question 240 A pipeline requires an approval before a generated change can become a release. Where should that gate live? In a local developer note inaccessible to CI. In a message asking the model to be careful. Only after production has been deployed. In CI/CD branch protection or deployment controls, not only in an assistant prompt. Explanation Correct Answer D:A real release permission must be enforced by the automation platform. Incorrect Answers A:The platform cannot enforce a rule it never sees. B:Prompt guidance cannot guarantee release authorization. C:Late approval does not prevent an unauthorized release.
Question 241 A user corrects an assistant’s earlier assumption about the permitted export audience. What should a long-session summary do? Treat the initial assistant assumption as higher priority than the user correction. Preserve the corrected audience constraint and mark the earlier assumption superseded. Keep both conflicting constraints as equally active. Discard the correction to shorten context. Explanation Correct Answer B:The authoritative newer instruction should control later actions. Incorrect Answers A:An assistant’s earlier guess is not authority. C:That creates ambiguous authorization. D:The summary would reintroduce a known error.
Question 242 An MCP search tool returns results with the wrong schema: ids are strings in one record and null in another. How should the agent gateway respond? Treat all results as valid because the tool name was correct. Validate the contract and surface an explicit inconsistent-output error. Delete the invalid record silently without disclosing missing data. Replace null identifiers with guessed sequential numbers. Explanation Correct Answer B:Consumers should not infer stable record identity from malformed upstream data. Incorrect Answers A:Tool identity is not evidence of response-contract compliance. C:Silent omission can conceal incomplete retrieval. D:Fabricated IDs may point to unrelated records.
Question 243 A company has 120,000 independent support tickets to classify overnight, with no interactive latency requirement. Which pattern is most suitable? Require a human to approve every transport request. Submit independent requests through an asynchronous batch workflow and reconcile results by custom ID. Discard per-ticket identifiers before submission. Keep a single interactive chat open for all tickets. Explanation Correct Answer B:Batching matches noninteractive throughput and asynchronous completion. Incorrect Answers A:the goal is scalable processing of routine independent items. C:Result association would become unreliable. D:That couples unrelated tickets into one context and runtime.
Question 244 A small change passes tests but formatting checks fail. What should the agent do? Delete the formatter configuration globally. Run deployment to prove the formatter is irrelevant. Ignore formatting because runtime tests passed. Apply the project’s formatting convention to the affected files and rerun the check. Explanation Correct Answer D:A focused style correction completes the actual requested quality gate. Incorrect Answers A:Removing the standard is not a targeted fix. B:Production release does not resolve a lint failure. C:The project may require both functional and formatting checks.
Question 245 A host runs an autonomous ticket agent in production. Which TWO controls make its tool loop both safe and auditable? Persist tool identifiers, arguments, and results with sensitive data appropriately redacted. Remove tool outcome messages to reduce token use in every run. Allow unrestricted calls indefinitely because the model usually finishes. Treat any HTTP success as proof that the user’s request was fulfilled. Set bounded runtime or call budgets with explicit stop and escalation behavior. Explanation Correct Answers A:A trace enables reconstruction of decisions and failures without exposing secrets. E:A limit contains runaway execution and gives a clear outcome when progress stalls. Incorrect Answers B:Omitting results can make subsequent decisions blind to important state. C:Normal completion rates do not remove the possibility of runaway or costly loops. D:Tool success can be only an intermediate step or the wrong action.
Question 246 A trust-and-safety model is accurate on most ordinary messages but uncertain on rare policy exceptions. What review threshold is suitable? Suppress all borderline cases without recording them. Treat numerical confidence alone as the whole policy. Escalate low-evidence, high-impact or policy-borderline cases with context and model uncertainty. Review only messages that the model already labels certain. Explanation Correct Answer C:Calibration should focus human attention on the costly failure modes. Incorrect Answers A:Invisible filtering impedes accountability. B:The consequences and type of uncertainty also matter. D:That ignores the cases most likely to need judgment.
Question 247 A support workflow retries a timed-out refund tool call. The first call may have succeeded despite the missing response. What design avoids duplicate refunds? Always submit a fresh refund with a new transaction ID. Treat a client timeout as proof no charge was reversed. Use a stable idempotency key and reconcile operation status before retrying a side effect. Remove the transaction audit trail to reduce confusion. Explanation Correct Answer C:An unknown outcome must be resolved against the transactional system. Incorrect Answers A:Fresh IDs defeat duplicate suppression. B:Response loss does not prove operation failure. D:That destroys the evidence needed to reconcile.
Question 248 A coordinator can delegate to three specialists but a code scanner is never relevant for policy questions. What should its tool list reflect? Disable the specialist tools that actually retrieve policy documents. Offer only delegation and retrieval capabilities needed for the active policy workflow. Use command execution as the default answer tool. Always register every possible scanner and mutator. Explanation Correct Answer B:Filtering irrelevant tools reduces choices and accidental actions. Incorrect Answers A:That removes needed evidence channels. C:Shell execution is not intrinsically required for policy research. D:Unrelated capabilities increase routing errors.
Question 249 A compliance team wants tool-level observability without breaking agent execution. Which TWO hook properties are important? Store every access token verbatim for debugging. Define clear behavior when the logging service fails, without silently granting extra permissions. Allow hooks to run arbitrary unapproved business actions. Record correlation identifiers and normalized outcome statuses for each intercepted call. Rewrite every tool outcome to success to keep dashboards green. Explanation Correct Answers B:Observability failure must not become an authorization bypass. D:Traces can tie activity back to a specific run and tool result. Incorrect Answers A:Raw tokens are dangerous unnecessary logging data. C:Hooks should not become uncontrolled privileged side-effect engines. E:False outcomes would mislead both operations and the agent.
Question 250 A team stores a rule for database migrations but a developer uses it while editing mobile UI strings. What is the likely rule-design flaw? The editor must delete all mobile files. The rule is too globally scoped and should match migration paths only. The rule needs a larger maximum token budget. All migration policies must apply to UI copy. Explanation Correct Answer B:Conditional context should not be loaded for unrelated UI work. Incorrect Answers A:Deleting UI code is unrelated to rule scoping. C:More tokens do not change its applicability. D:The targets have different requirements.
