Question 180 A user changes the request from “propose a patch” to “apply the approved patch” in the same workspace. What must be updated? The execution authorization and tool mode, while preserving the vetted plan. Permission to modify any other repository. A random new goal unrelated to the patch. The entire project identity and all file IDs. Explanation Correct Answer A:A clear approval can justify a controlled transition from planning to editing. Incorrect Answers B:Approval for one patch is not universal write authority. C:The approved scope should remain the objective. D:Mode changes do not require restarting the project.
Question 181 A support conversation includes a verified account number from six turns earlier. After compaction, the agent loses it. What should be retained explicitly? A fresh guessed account number from the model. Only a generic label such as “customer issue.” A durable fact record of verified identifiers and relevant constraints across compaction. The assistant’s last thank-you message. Explanation Correct Answer C:Compaction should preserve operationally important values that later decisions depend on. Incorrect Answers A:Guessing an identifier can lead to cross-account actions. B:The verified identity cannot be reconstructed from that summary. D:Courtesy text does not preserve the critical account state.
Question 182 A model repeatedly answers an invoice question from memory even though current invoice data must be fetched. What tool-choice strategy could require a retrieval step? Increase output creativity until the model happens to call a tool. Grant the agent permission to issue refunds instead. Disable the invoice tool and ask for a confident estimate. Constrain the relevant invocation to use the approved invoice-read tool, then permit synthesis. Explanation Correct Answer D:A controlled tool choice can enforce evidence acquisition where required. Incorrect Answers A:Randomness is not a reliable enforcement method. B:A financial write action does not retrieve an invoice. C:That makes current-state verification impossible.
Question 183 A tool-calling assistant mistakes order identifiers for invoice identifiers. Which demonstration has the highest value? A case with only one ID type visible and no ambiguity. An unrelated weather lookup example. A sample response that claims IDs are interchangeable. An example where both identifiers appear and the correct API parameter is chosen explicitly. Explanation Correct Answer D:A realistic contrast teaches disambiguation at the tool boundary. Incorrect Answers A:The failure occurs when competing identifiers coexist. B:It does not address business identifier selection. C:Wrong equivalence would cause misrouting.
Question 184 A research agent must resume after context compaction. Which information is most important to retain? A fresh search plan that ignores all completed research. Every greeting and unrelated conversational aside in chronological order. Verified source links, open questions, completed evidence checks and the exact next unresolved task. An unsourced final conclusion without the underlying citations. Explanation Correct Answer C:This prevents redoing work or treating unresolved facts as settled. Incorrect Answers A:Restarting wastes work and can repeat mistakes. B:Low-value dialogue crowds out the operational checkpoint. D:The resumed agent cannot distinguish evidence from inference.
Question 185 An agent’s first code patch fails type checking due to one nullability mismatch. What is a sensible loop? Inspect the compiler diagnostic, update the relevant type handling, and rerun affected checks. Declare success because the first patch was syntactically valid. Regenerate the entire codebase from scratch. Suppress all compiler errors globally. Explanation Correct Answer A:Specific feedback guides a bounded correction and verifies the result. Incorrect Answers B:Type checking still failed and must be addressed. C:This is disproportionate to a local type mismatch. D:Removing checks does not make types safe.
Question 186 A worker finds a critical missing document while another branch is already drafting. How should the coordinator manage the discovery? Terminate all agents and delete completed work. Flag the dependency, request the missing evidence, and defer claims that require it. Ask the drafter to invent the absent document. Hide the dependency because the drafting task has started. Explanation Correct Answer B:A subagent discovery can change the evidence required for downstream synthesis. Incorrect Answers A:Only affected claims need to wait for the new evidence. C:A fabricated source is not a valid substitution. D:Publishing unsupported claims would turn a recoverable gap into a factual error.
Question 187 A code-generation agent produces a security-sensitive patch. What should a separate reviewer inspect? Whether both agents use identical wording. Only the first agent’s self-rated confidence. Only the prose summary without the patch. The changed diff, relevant threat model and tests against explicit invariants. Explanation Correct Answer D:A second pass must verify behavior, not merely endorse a fluent explanation. Incorrect Answers A:Matching language is not independent verification. B:Confidence alone does not establish correctness. C:That misses implementation errors.
Question 188 A tool has an optional dry_run parameter but its behavior is undocumented. A finance agent executes real transactions during tests. What should the description state? Assume every call is a simulation. Explicitly describe whether dry_run suppresses side effects and make safe testing defaults enforceable. Depend on an arbitrary naming convention in conversation history. Hide dry_run from the schema so it cannot be used. Explanation Correct Answer B:Models and clients need an unambiguous sandbox versus live-action distinction. Incorrect Answers A:An undocumented assumption can move real funds. C:Runtime behavior should be specified in the tool contract. D:That removes the available safety mechanism.
Question 189 A review queue handles AI-generated incident reports. Which TWO controls most improve qualified human decisions? Require acknowledgement before any authorized external notification is sent. Treat a high confidence score as proof the incident occurred. Show the supporting logs, source timestamps and explicit unresolved assumptions. Hide the original evidence to prevent independent challenge. Let the model mark human approval as complete automatically. Explanation Correct Answers A:A consequential communication should follow the defined approval policy. C:Reviewers need evidence and uncertainty rather than a fluent summary alone. Incorrect Answers B:Confidence labels do not establish event truth. D:That prevents meaningful verification. E:The approving party must actually act.
Question 190 A skill is intended to read repository files but not execute shell commands. Which metadata is most relevant? A promise that the model will never be tempted to run Bash. A request for unrestricted production credentials. An allowed-tools declaration limited to read-oriented capabilities. An arbitrary font name in the skill file. Explanation Correct Answer C:Tool scoping lets the skill operate with least privilege. Incorrect Answers A:A prompt promise is weaker than an enforceable capability boundary. B:The read task does not require them. D:Presentation choices do not constrain tool use.
