Question 610
The release checklist for a customer case-management agent explicitly checks that out-of-policy exceptions receive an authorized decision; it also checks that each tool interface has a single clear operation boundary. What coordinated solution meets the checklist?
- let the agent approve an exception whenever its wording sounds confident; and separate read-only lookup from side-effecting update interfaces
- route the exception to a human approval step before execution; and use a single universal tool that infers read versus write from prose
- let the agent approve an exception whenever its wording sounds confident; and use a single universal tool that infers read versus write from prose
- route the exception to a human approval step before execution; and separate read-only lookup from side-effecting update interfaces
Explanation
Correct Answer
D:In a customer case-management agent, approval is a workflow boundary separate from language-model confidence; alongside this, clear operation boundaries support least privilege and understandable approvals, satisfying the distinct requirements that out-of-policy exceptions receive an authorized decision and that each tool interface has a single clear operation boundary.
Incorrect Answers
A:In a customer case-management agent, clear operation boundaries support least privilege and understandable approvals, but model confidence is not delegated business authority; therefore the separate requirement that out-of-policy exceptions receive an authorized decision would remain unsatisfied.
B:In a customer case-management agent, approval is a workflow boundary separate from language-model confidence, but overloading hides effects and makes execution review harder; therefore the separate requirement that each tool interface has a single clear operation boundary would remain unsatisfied.
C:In a customer case-management agent, model confidence is not delegated business authority and overloading hides effects and makes execution review harder; the design would fail both conditions: out-of-policy exceptions receive an authorized decision and each tool interface has a single clear operation boundary.
Question 611
An architecture review of a claims form processing assistant finds two missing guarantees: a model learns a boundary between superficially similar cases; thousands of independent documents can finish without interactive response deadlines. What two-part change closes both gaps?
- include contrasting positive and negative demonstrations; and use asynchronous Message Batches processing for the bulk workload
- include contrasting positive and negative demonstrations; and make every document a synchronous request inside the user interaction
- show only easy positive samples; and use asynchronous Message Batches processing for the bulk workload
- show only easy positive samples; and make every document a synchronous request inside the user interaction
Explanation
Correct Answer
A:In a claims form processing assistant, near-boundary examples communicate the decisive distinction and reduce false positives; alongside this, batching suits non-blocking independent requests and can reduce processing cost, satisfying the distinct requirements that a model learns a boundary between superficially similar cases and that thousands of independent documents can finish without interactive response deadlines.
Incorrect Answers
B:In a claims form processing assistant, near-boundary examples communicate the decisive distinction and reduce false positives, but serial immediate calls add latency where no caller is waiting; therefore the separate requirement that thousands of independent documents can finish without interactive response deadlines would remain unsatisfied.
C:In a claims form processing assistant, batching suits non-blocking independent requests and can reduce processing cost, but the model may generalize the label to unrelated but similar-looking items; therefore the separate requirement that a model learns a boundary between superficially similar cases would remain unsatisfied.
D:In a claims form processing assistant, the model may generalize the label to unrelated but similar-looking items and serial immediate calls add latency where no caller is waiting; the design would fail both conditions: a model learns a boundary between superficially similar cases and thousands of independent documents can finish without interactive response deadlines.
Question 612
Tests for a security-advisory aggregation workflow will pass only if a parser preserves the specified output structure across cases and if the agent can recover exact policy facts after compaction. Which proposed design directly satisfies both checks?
- mix contradictory schemas in the example set; and treat the compressed summary as an immutable source of truth
- keep each demonstration's field order and types consistent; and retrieve authoritative facts from a durable store when needed
- keep each demonstration's field order and types consistent; and treat the compressed summary as an immutable source of truth
- mix contradictory schemas in the example set; and retrieve authoritative facts from a durable store when needed
Explanation
Correct Answer
B:In a security-advisory aggregation workflow, coherent exemplars teach a stable mapping rather than competing formats; alongside this, re-reading source state restores precision that a condensed history cannot guarantee, satisfying the distinct requirements that a parser preserves the specified output structure across cases and that the agent can recover exact policy facts after compaction.
Incorrect Answers
A:In a security-advisory aggregation workflow, inconsistent samples can make expected output ambiguous and summaries may simplify caveats or omit an exception; the design would fail both conditions: a parser preserves the specified output structure across cases and the agent can recover exact policy facts after compaction.
C:In a security-advisory aggregation workflow, coherent exemplars teach a stable mapping rather than competing formats, but summaries may simplify caveats or omit an exception; therefore the separate requirement that the agent can recover exact policy facts after compaction would remain unsatisfied.
D:In a security-advisory aggregation workflow, re-reading source state restores precision that a condensed history cannot guarantee, but inconsistent samples can make expected output ambiguous; therefore the separate requirement that a parser preserves the specified output structure across cases would remain unsatisfied.
Question 613
For a procurement research assistant, a reviewer identifies two independent obligations: one specialist's crash does not discard other verified outputs and a risk prompt handles missing supporting evidence honestly. Which proposal should be approved?
- isolate branch failures and preserve completed artifacts for coordination; and instruct the model to return unknown or needs-review when evidence is absent
- restart every agent and erase prior results after a single crash; and require a definitive yes-or-no answer for every incomplete record
- restart every agent and erase prior results after a single crash; and instruct the model to return unknown or needs-review when evidence is absent
- isolate branch failures and preserve completed artifacts for coordination; and require a definitive yes-or-no answer for every incomplete record
Explanation
Correct Answer
A:In a procurement research assistant, fault containment avoids losing good work while a failed branch is repaired; alongside this, a defined abstention path avoids forced confident judgments without facts, satisfying the distinct requirements that one specialist's crash does not discard other verified outputs and that a risk prompt handles missing supporting evidence honestly.
Incorrect Answers
B:In a procurement research assistant, unaffected branches would be redone and provenance could be lost and forced binaries hide uncertainty as false certainty; the design would fail both conditions: one specialist's crash does not discard other verified outputs and a risk prompt handles missing supporting evidence honestly.
C:In a procurement research assistant, a defined abstention path avoids forced confident judgments without facts, but unaffected branches would be redone and provenance could be lost; therefore the separate requirement that one specialist's crash does not discard other verified outputs would remain unsatisfied.
D:In a procurement research assistant, fault containment avoids losing good work while a failed branch is repaired, but forced binaries hide uncertainty as false certainty; therefore the separate requirement that a risk prompt handles missing supporting evidence honestly would remain unsatisfied.
Question 614
An architecture review of a security-advisory aggregation workflow finds two missing guarantees: a truncated response is not mistaken for successful completion; a required validation step runs before the final response even when the model does not support forced tool choice. What two-part change closes both gaps?
- handle max_tokens as truncation rather than end_turn; and rely on a prompt asking the model to call the validator
- handle max_tokens as truncation rather than end_turn; and enforce the validator in application control flow, using forced tool_choice only with compatible models
- interpret every non-tool stop reason as a successful final answer; and rely on a prompt asking the model to call the validator
- implement only handle max_tokens as truncation rather than end_turn; make no corresponding change for the second requirement
- interpret every non-tool stop reason as a successful final answer; and enforce the validator in application control flow, using forced tool_choice only with compatible models
Explanation
Correct Answer
B:In a security-advisory aggregation workflow, reaching an output limit is a stop condition but does not establish task completion; alongside this, application-level orchestration guarantees the required stage across models, while forced tool_choice is limited to compatible models and settings, satisfying the distinct requirements that a truncated response is not mistaken for successful completion and that a required validation step runs before the final response even when the model does not support forced tool choice.
Incorrect Answers
A:In a security-advisory aggregation workflow, reaching an output limit is a stop condition but does not establish task completion, but prompt instructions do not guarantee execution of the validator or override model-specific tool-choice restrictions; therefore the separate requirement that a required validation step runs before the final response even when the model does not support forced tool choice would remain unsatisfied.
C:In a security-advisory aggregation workflow, a length-limited result may be incomplete or unusable and prompt instructions do not guarantee execution of the validator or override model-specific tool-choice restrictions; the design would fail both conditions: a truncated response is not mistaken for successful completion and a required validation step runs before the final response even when the model does not support forced tool choice.
D:In a security-advisory aggregation workflow, reaching an output limit is a stop condition but does not establish task completion, but the independent mechanism for the other criterion is absent: application-level orchestration guarantees the required stage across models, while forced tool_choice is limited to compatible models and settings; as written, the design does not ensure that a required validation step runs before the final response even when the model does not support forced tool choice.
E:In a security-advisory aggregation workflow, application-level orchestration guarantees the required stage across models, while forced tool_choice is limited to compatible models and settings, but a length-limited result may be incomplete or unusable; therefore the separate requirement that a truncated response is not mistaken for successful completion would remain unsatisfied.
Question 615
Before expanding a customer refund workflow, the owner must demonstrate that quality drift is detected after a workflow launches, while also demonstrating that conflicting sources are not silently flattened into one answer. Which design provides both controls?
- sample outcomes for periodic human QA against known criteria; and report the disagreement and explain evidence quality or recency
- validate one development example and never review production output; and report the disagreement and explain evidence quality or recency
- sample outcomes for periodic human QA against known criteria; and choose one assertion and omit the contradiction without investigation
- validate one development example and never review production output; and choose one assertion and omit the contradiction without investigation
Explanation
Correct Answer
A:In a customer refund workflow, ongoing calibration can reveal changed input distributions and silent error patterns; alongside this, explicit conflicts preserve uncertainty rather than manufacturing unanimity, satisfying the distinct requirements that quality drift is detected after a workflow launches and that conflicting sources are not silently flattened into one answer.
Incorrect Answers
B:In a customer refund workflow, explicit conflicts preserve uncertainty rather than manufacturing unanimity, but production inputs may differ from the initial examples; therefore the separate requirement that quality drift is detected after a workflow launches would remain unsatisfied.
C:In a customer refund workflow, ongoing calibration can reveal changed input distributions and silent error patterns, but suppressed disagreement creates false confidence; therefore the separate requirement that conflicting sources are not silently flattened into one answer would remain unsatisfied.
D:In a customer refund workflow, production inputs may differ from the initial examples and suppressed disagreement creates false confidence; the design would fail both conditions: quality drift is detected after a workflow launches and conflicting sources are not silently flattened into one answer.
Question 616
For a pre-merge code inspection workflow, acceptance depends on demonstrating that a high-risk workflow only invokes specifically approved capabilities. Which approach is the strongest fit?
- rely on a sentence asking the agent not to use risky tools
- create a documented reusable skill in SKILL.md
- declare allowed-tools for the skill and enforce executor permissions
- define how the skill accepts and uses invocation arguments
Explanation
Correct Answer
C:For a pre-merge code inspection workflow, tool restrictions narrow the agent's operational capability while the skill runs; this directly establishes that a high-risk workflow only invokes specifically approved capabilities instead of relying on an unrelated workflow control.
Incorrect Answers
A:In a pre-merge code inspection workflow, instruction-only boundaries cannot substitute for actual tool authorization; as a result, the stated requirement that a high-risk workflow only invokes specifically approved capabilities would still not be met.
B:Although skills package instructions and supporting resources for invocation across similar tasks, that only supports the distinct goal that a repeatable code-review procedure is available as a named workflow, not the requirement that a high-risk workflow only invokes specifically approved capabilities in a pre-merge code inspection workflow.
D:Although parameterized workflows apply one reviewed procedure to different resources without hardcoding targets, that only supports the distinct goal that a shared skill can take a supplied target name or file path, not the requirement that a high-risk workflow only invokes specifically approved capabilities in a pre-merge code inspection workflow.
Question 617
For a repository documentation project, a reviewer identifies two independent obligations: two teams' analyses are reconciled before publication and the agent executes tests and examines their exit status. Which proposal should be approved?
- add a synthesis stage with explicit conflict-resolution criteria; and use Bash with an approved test command
- publish whichever branch finishes last; and inspect the test configuration only with Read and claim tests passed
- add a synthesis stage with explicit conflict-resolution criteria; and inspect the test configuration only with Read and claim tests passed
- publish whichever branch finishes last; and use Bash with an approved test command
Explanation
Correct Answer
A:In a repository documentation project, merging divergent work requires an authoritative step rather than choosing by arrival time; alongside this, the shell can run build tools and return execution feedback that Read cannot supply, satisfying the distinct requirements that two teams' analyses are reconciled before publication and that the agent executes tests and examines their exit status.
Incorrect Answers
B:In a repository documentation project, run order is not an evidence or quality standard and reading configuration does not execute a test suite; the design would fail both conditions: two teams' analyses are reconciled before publication and the agent executes tests and examines their exit status.
C:In a repository documentation project, merging divergent work requires an authoritative step rather than choosing by arrival time, but reading configuration does not execute a test suite; therefore the separate requirement that the agent executes tests and examines their exit status would remain unsatisfied.
D:In a repository documentation project, the shell can run build tools and return execution feedback that Read cannot supply, but run order is not an evidence or quality standard; therefore the separate requirement that two teams' analyses are reconciled before publication would remain unsatisfied.
Question 618
Before expanding a financial statement data extractor, the owner must demonstrate that sensitive tool calls are denied before they reach an executor, while also demonstrating that a broad tool catalog does not overwhelm a small task. Which design provides both controls?
- write a stronger instruction in the system prompt; and supply only context-relevant tools for the current agent
- write a stronger instruction in the system prompt; and load all available MCP tools into every specialist turn
- register a PreToolUse hook that checks the operation against policy; and supply only context-relevant tools for the current agent
- register a PreToolUse hook that checks the operation against policy; and load all available MCP tools into every specialist turn
Explanation
Correct Answer
C:In a financial statement data extractor, pre-execution interception can reject a disallowed call rather than relying on post-hoc advice; alongside this, reducing irrelevant tool definitions improves selection and saves context, satisfying the distinct requirements that sensitive tool calls are denied before they reach an executor and that a broad tool catalog does not overwhelm a small task.
Incorrect Answers
A:In a financial statement data extractor, reducing irrelevant tool definitions improves selection and saves context, but an instruction alone does not provide deterministic execution-time authorization; therefore the separate requirement that sensitive tool calls are denied before they reach an executor would remain unsatisfied.
B:In a financial statement data extractor, an instruction alone does not provide deterministic execution-time authorization and tool bloat makes the available actions harder to discriminate; the design would fail both conditions: sensitive tool calls are denied before they reach an executor and a broad tool catalog does not overwhelm a small task.
D:In a financial statement data extractor, pre-execution interception can reject a disallowed call rather than relying on post-hoc advice, but tool bloat makes the available actions harder to discriminate; therefore the separate requirement that a broad tool catalog does not overwhelm a small task would remain unsatisfied.
Question 619
Production readiness for a dependency upgrade initiative depends on ensuring that delegated investigations stay within bounded scope and that a broad tool catalog does not overwhelm a small task. Select the most appropriate combined approach.
- ask every subagent to investigate everything it can find; and load all available MCP tools into every specialist turn
- ask every subagent to investigate everything it can find; and supply only context-relevant tools for the current agent
- specify the question, allowed sources, expected output, and stopping criterion; and load all available MCP tools into every specialist turn
- specify the question, allowed sources, expected output, and stopping criterion; and supply only context-relevant tools for the current agent
Explanation
Correct Answer
D:In a dependency upgrade initiative, a task contract prevents specialist exploration from swallowing the coordinator's larger goal; alongside this, reducing irrelevant tool definitions improves selection and saves context, satisfying the distinct requirements that delegated investigations stay within bounded scope and that a broad tool catalog does not overwhelm a small task.
Incorrect Answers
A:In a dependency upgrade initiative, unbounded exploration dilutes signal and inflates context and latency and tool bloat makes the available actions harder to discriminate; the design would fail both conditions: delegated investigations stay within bounded scope and a broad tool catalog does not overwhelm a small task.
B:In a dependency upgrade initiative, reducing irrelevant tool definitions improves selection and saves context, but unbounded exploration dilutes signal and inflates context and latency; therefore the separate requirement that delegated investigations stay within bounded scope would remain unsatisfied.
C:In a dependency upgrade initiative, a task contract prevents specialist exploration from swallowing the coordinator's larger goal, but tool bloat makes the available actions harder to discriminate; therefore the separate requirement that a broad tool catalog does not overwhelm a small task would remain unsatisfied.
Question 620
An audit of a repository maintenance copilot asks how to make sure that conflicting subagent findings produce one accountable interpretation. Which implementation should be selected?
- let the fastest subagent override other findings automatically
- have the coordinator compare claims and sources before synthesis
- record per-branch status and let the coordinator choose retry or partial response
- give subagents a common results contract including claim, evidence, and status
- run retrieval before analysis and pass its validated artifacts
Explanation
Correct Answer
B:For a repository maintenance copilot, central reconciliation can distinguish evidence disagreement from a formatting difference; this directly establishes that conflicting subagent findings produce one accountable interpretation instead of relying on an unrelated workflow control.
Incorrect Answers
A:In a repository maintenance copilot, completion order is not a sound evidence-ranking rule; as a result, the stated requirement that conflicting subagent findings produce one accountable interpretation would still not be met.
C:Although the orchestration layer must distinguish missing evidence from a negative factual finding, that only supports the distinct goal that the system can continue when one specialist returns an explicit failure, not the requirement that conflicting subagent findings produce one accountable interpretation in a repository maintenance copilot.
D:Although a typed handoff lets the coordinator compare like fields and detect missing support, that only supports the distinct goal that specialists return comparable artifacts for aggregation, not the requirement that conflicting subagent findings produce one accountable interpretation in a repository maintenance copilot.
E:Although a downstream agent needs actual upstream findings rather than an assumed future result, that only supports the distinct goal that dependent analysis begins only when source retrieval is complete, not the requirement that conflicting subagent findings produce one accountable interpretation in a repository maintenance copilot.