Question 370 An insurer needs documents translated, extracted, and checked for policy coverage. Some files are already translated. Which approach avoids waste? Check artifact state and execute only missing, dependency-ordered steps for each document. Skip extraction when translation exists. Run every stage concurrently even when one requires the previous output. Retranslate every file whenever any stage changes. Explanation Correct Answer A:Workflow routing can reuse completed translations while preserving required extraction and review. Incorrect Answers B:Translation does not create the coverage facts the next stage needs. C:Dependency violations can produce invalid or empty results. D:This repeats completed work and may introduce inconsistencies.
Question 371 A reviewer calls a function insecure solely because it uses a deprecated helper name, although the helper is still supported in this pinned version. What should it check? All functions with old-fashioned names should fail automatically. Only the number of functions in the file. The project’s actual dependency version and evidence of a relevant unsafe behavior. A random current version from the internet without inspecting the lockfile. Explanation Correct Answer C:Version-scoped criteria avoid assuming every deprecated-looking name is a security bug. Incorrect Answers A:Names do not prove exploitation. B:File size is unrelated to the disputed helper behavior. D:Project-specific dependencies determine applicable API behavior.
Question 372 A claims agent must check policy eligibility before issuing a settlement. Testing shows it occasionally pays first and checks later. Which design closes the gap? Add three examples of the correct sequence to the prompt. Review only the end-of-day list of unauthorized settlements. Require an eligibility decision token at the settlement-tool boundary. Reduce the maximum tool count to two. Explanation Correct Answer C:The payment service cannot execute until an independently checked prerequisite succeeds. Incorrect Answers A:Examples can improve average behavior but do not enforce a mandatory order. B:Detection after disbursement does not prevent the business violation. D:An arbitrary call cap does not guarantee eligibility precedes settlement.
Question 373 A debugging agent modifies a function and reports “fixed” without running the regression test the user specified. What is missing? A lower randomness setting for its final response. A commit message written before testing. An executed regression check with a recorded result for the changed code. A longer justification of why the patch seems elegant. Explanation Correct Answer C:Source edits alone cannot substantiate a behavior claim. Incorrect Answers A:Sampling controls cannot prove a defect is fixed. B:Commit notes cannot replace execution evidence. D:Prose does not establish runtime success.
Question 374 A research agent proposes three independent repository searches in one response. What is the proper client-side handling? Combine all search results under the first tool-use identifier. Run only the first search because a message can contain just one actionable tool. Run all permitted tool requests and return one matching result for each tool-use identifier. Wait for an end_turn before running any of the three searches. Explanation Correct Answer C:Each use block needs its own result, even if searches can execute concurrently. Incorrect Answers A:Mismatched result identifiers corrupt the correspondence between requests and results. B:A single assistant message can request multiple client tools. D:tool_use is already the signal to execute the requested operations.
Question 375 A retry loop repeatedly reformats an answer but keeps inventing nonexistent citations. What termination rule should apply? After a bounded correction attempt, stop and escalate missing evidence rather than manufacturing citations. Suppress all citations without noting the missing source. Use URLs from unrelated records as replacements. Keep retrying until a citation string looks valid. Explanation Correct Answer A:A semantic evidence gap cannot be solved by infinite formatting retries. Incorrect Answers B:That conceals an evidence limitation. C:Provenance must match the claim. D:An appealing citation may still be fabricated.
Question 376 A support case has separate billing and shipping concerns. Both require the same verified customer identity. How should work be divided? Ask the final response generator to reconstruct identity from partial records. Run both branches before verifying whose account is involved. Require shipping investigation to finish before billing despite no dependency. Verify identity once, then run independent billing and shipping investigations in parallel. Explanation Correct Answer D:Shared prerequisites precede safe parallel branches that do not depend on one another. Incorrect Answers A:Synthesis cannot replace a required identity check. B:That risks using unauthorized or incorrect customer records. C:Artificial serial order increases latency.
Question 377 A user approves a narrowly scoped typo fix and asks Claude Code to apply it. When is direct execution reasonable? By disabling all version-control checks permanently. Only after building an elaborate multi-agent architecture document. When the requested change and authorization are clear, with appropriate validation afterward. Without inspecting the target file at all. Explanation Correct Answer C:A small bounded edit need not require another lengthy planning-only round. Incorrect Answers A:Directness does not remove the need for safe editing. B:Overplanning simple authorized changes adds unnecessary cost. D:Direct execution still needs to ground the edit in current content.
Question 378 An investigation goal spans malware, finance, and privacy; every worker gets a vague one-line instruction and produces different output formats. What should the coordinator specify? A rule that specialists cannot report uncertainty. Only a final word count for each response. One generic prompt copied to every worker regardless of expertise. Distinct scoped assignments plus a shared evidence schema for handoff. Explanation Correct Answer D:Clear specialist contracts make heterogeneous results comparable and auditable. Incorrect Answers A:Hidden uncertainty undermines synthesis and exception handling. B:Output length cannot specify the evidence fields downstream workers require. C:Identical vague instructions miss domain-specific responsibilities.
Question 379 An architect needs line-level evidence in structured extraction output. What field design is appropriate? Replace evidence with a confidence percentage generated by the model. Include the extracted value together with source document ID and location or span. Copy the entire unrelated source corpus into every record. Return only values with no source mappings. Explanation Correct Answer B:Typed provenance fields allow auditors to verify each value. Incorrect Answers A:A confidence label does not locate the claimed fact. C:Excess context burdens storage and may expose unnecessary data. D:The downstream team cannot distinguish extraction from invention.
Question 380 A research request asks for “all major causes” of a system outage; four agents independently research network failures and none studies storage or identity. Which root cause is likely? Too little output length in the network findings. The final writer lacks a decorative summary heading. Insufficient API temperature to generate varied prose. Poor task partitioning with duplicated scope and omitted categories. Explanation Correct Answer D:The work split fails both coverage and non-overlap requirements. Incorrect Answers A:Longer network reports cannot reveal unassigned categories. B:Formatting does not restore missing investigations. C:Prose variation is unrelated to systematic task coverage.
