Question 190
A support response must avoid promising an action until the API confirms it. What should the instructions say?
- State required positive findings, limitations, risks and outstanding decisions
- Require tool-confirmed completion before wording any definitive customer promise
- Define explicit eligibility conditions and required evidence for each refund decision
- Assign decisions and proposals to named sources, preserving unresolved disagreement
Explanation
Correct Answer
B:The prompt can distinguish proposed, attempted and confirmed actions. This prevents the assistant from inferring financial success from its own intent.
Incorrect Answers
A:Because balanced section requirements make omissions more visible during validation, the measure supports eliciting both benefits and risks in reports rather than the needed preventing premature promises of side effects.
C:Concrete criteria make a model's judgment testable against policy; supporting translating broad intent into checkable refund rules cannot replace the required preventing premature promises of side effects.
D:Attribution is part of factual correctness; this contributes to preserving attribution in generated meeting summaries without meeting the need for preventing premature promises of side effects.
Question 191
An invoice extractor confuses purchase-order numbers with invoice numbers despite detailed field definitions. What prompting technique may help?
- Use layout-diverse examples but never label which nearby field is the invoice number versus the purchase-order number
- Validate on held-out realistic tickets rather than judging only the shown demonstrations
- Present diverse input date formats paired with one required normalized representation
- Provide a few labeled input-output examples illustrating the distinction
- Employ varied fictional entities and emphasize that examples illustrate structure, not actual values
Explanation
Correct Answer
D:Representative examples demonstrate how to resolve ambiguous document fields. They complement rather than replace the field schema and validation checks.
Incorrect Answers
A:Changing layouts without labeling the commonly confused identifiers does not teach the distinction. Worked labeled examples must demonstrate how to separate PO and invoice numbers.
B:For checking generalization beyond few-shot examples, in-prompt examples are teaching material, not an independent evaluation set; however, this requirement concerns teaching field disambiguation by example.
C:Contrasting examples demonstrate conversion rules while preserving a single target shape; that may serve demonstrating cross-locale date normalization, but it is no substitute for teaching field disambiguation by example.
E:Reusing the same names across all demonstrations can invite copying; the stated problem requires teaching field disambiguation by example rather than preventing sample values from leaking into outputs.
Question 192
A developer knows a filename fragment but not the file contents. Which builtin search avoids expensive text scanning?
- Employ Edit for a targeted modification to an existing file
- Employ Read to examine the specified file
- Use Glob for pathname discovery before reading content
- Run a Grep search for occurrences, inspect affected files, then apply controlled Edits
Explanation
Correct Answer
C:Path globbing matches file names without requiring content search. Grep would answer a different question about occurrences of text inside files.
Incorrect Answers
A:Targeted editing changes selected text while maintaining the surrounding source; the deciding condition is searching by path rather than file contents, whereas this action concerns changing a limited region of existing code.
B:Read provides file content for reasoning without modifying the repository; the option would serve inspecting a known file without edits, not searching by path rather than file contents as specified in the stem.
D:A search-first workflow reveals scope and avoids blind global replacement; that relates to scoping a multi-file refactor before mutation, not the necessary searching by path rather than file contents.
Question 193
An agent replaces a small function with a larger abstraction that increases maintenance burden. What review standard is useful?
- Begin by analyzing the root cause before attempting another identical patch
- Retain changes bounded to the requested behavior and a small verified diff
- Prefer the smallest clear change that meets requirements and preserves interfaces
- Establish behavioral acceptance cases rather than treating lint success as completion
Explanation
Correct Answer
C:Simplicity is an engineering constraint: extra layers need a demonstrated benefit. A focused solution is generally easier to audit than speculative abstraction.
Incorrect Answers
A:When addressing escaping a repetitive unsuccessful repair loop, repeated unmodified retries do not add evidence; the present task instead calls for reviewing unnecessary complexity in generated code.
B:Narrow modifications are easier to review, test and revert; limiting a patch to the requested behavior and reviewing unnecessary complexity in generated code require different controls in this situation.
D:Linting checks style and some static issues; it does not prove runtime correctness; in this scenario, distinguishing static hygiene from functional success is secondary to reviewing unnecessary complexity in generated code.
Question 194
A customer case has two branches that must both pass before final resolution. Which task graph shape matches this?
- Collect closely related low-cost tasks into a meaningful worker assignment
- Give one bounded comparison to each worker and merge on a shared schema
- Adopt direct deterministic lookup and rule evaluation for the simple stages
- Create parallel prerequisite branches with a join gate before completion
Explanation
Correct Answer
D:A fan-out and join pattern preserves concurrency while enforcing that no required verification branch is skipped. A first-result-wins merge would be unsafe.
Incorrect Answers
A:Excessive decomposition multiplies setup and context cost; this can serve avoiding overfragmentation of delegated work, but leaves the requirement for joining mandatory parallel checks unmet.
B:Independent subtasks can run concurrently when they have comparable output contracts; that addresses decomposing parallel independent analyses, whereas this case calls for joining mandatory parallel checks.
C:Not every step requires another model or subagent; this concerns separating deterministic steps from inference, leaving the requirement for joining mandatory parallel checks outstanding.
Question 195
A product lookup returns no matching item for a valid SKU. What response semantics best reflect the outcome?
- Distinguish a legitimate not-found result from a tool execution error
- Label partial completion explicitly and identify which data remains unavailable
- Apply bounded retry-after guidance in the retry scheduler
- Normalize structured error codes and retry hints across tools
Explanation
Correct Answer
A:A successful search with zero matches is not necessarily an infrastructure failure. Treating it separately supports user clarification rather than retries against a healthy service.
Incorrect Answers
B:A partial response must not masquerade as complete evidence; this contributes to communicating incomplete tool output without meeting the need for separating no-match outcomes from failures.
C:Rate-limit responses can indicate when another attempt is reasonable; that serves honoring rate-limit backoff information, whereas the case requires separating no-match outcomes from failures.
D:Because consistent machine-readable categories make orchestration decisions repeatable, the measure supports normalizing tool error classifications rather than the needed separating no-match outcomes from failures.
Question 196
The design needs a list of discoverable operations from an MCP server rather than arbitrary prose instructions. What mechanism should it use?
- Rely on MCP tool discovery and declared tool interfaces
- Synchronize server identifier and tool-name changes with client configuration
- Adopt supported authenticated transport and credential renewal rather than embedding a fixed secret
- Confirm tool exposure, permission filters and server-side authorization
Explanation
Correct Answer
A:MCP presents structured capabilities with names, descriptions and schemas. This lets clients select operations using a defined protocol rather than inventing an untyped command channel.
Incorrect Answers
B:Clients refer to discovered capabilities using identifiers; this may support migrating MCP tool identifiers safely but cannot alone meet the requirement for discovering MCP tools through protocol metadata.
C:Remote MCP services need a secure authentication lifecycle; the stated problem requires discovering MCP tools through protocol metadata rather than maintaining authenticated MCP transport.
D:Connectivity alone does not grant every capability; a method suited to diagnosing a connected but unavailable MCP tool does not meet the need for discovering MCP tools through protocol metadata.
Question 197
A classification model reports 99% confidence but lacks supporting evidence for a fraud accusation. How should the reviewer treat this?
- Carry forward individual rationales, evidence and the final adjudication basis
- Employ automatic acceptance for verified low-risk cases and exception-based human review
- Order reviews by material impact, uncertainty and business risk
- Require evidence and calibration rather than trusting the self-reported confidence number
Explanation
Correct Answer
D:A model's expressed confidence may not correspond to measured probability. Important accusations need independent data and a reviewed decision process.
Incorrect Answers
A:Disagreement may expose ambiguous policies or missing facts; this addresses retaining review disagreement and adjudication provenance, while the question is about avoiding uncalibrated confidence as proof.
B:Hybrid workflows reserve human effort for uncertain or high-impact outcomes; that relates to balancing automation with targeted human oversight, not the necessary avoiding uncalibrated confidence as proof.
C:Although triage uses the cost of errors to allocate scarce review capacity, the choice addresses directing human attention to the highest-risk cases and not the requested avoiding uncalibrated confidence as proof.
Question 198
Two versions of a back-end tool return different keys for the same shipment status. What integration discipline helps?
- Record a stable idempotency key for mutation requests
- Publish a stable output schema and version incompatible changes
- Record search exclusions and filter parameters as explicit tool behavior
- Produce structured, bounded results with pagination and relevant fields
Explanation
Correct Answer
B:A predictable contract lets agents and validators consume results consistently. Silent field renames break downstream extraction even if the underlying API call succeeds.
Incorrect Answers
A:A client-provided operation identifier lets the service recognize repeat submissions and return the prior outcome; the unsatisfied requirement is managing output schema compatibility, not protecting a mutating tool against replay.
C:When considering revealing retrieval coverage limitations, the model can plan around a tool only when it knows the search domain; the requirement for managing output schema compatibility remains unmet.
D:When addressing controlling the volume of tool output, typed and paginated results let the agent request evidence progressively; the present task instead calls for managing output schema compatibility.
Question 199
A pricing agent can invoke other agents indefinitely to debate a small discount. Which architecture limits fan-out?
- Break up independent expert analyses and reserve integration for the orchestrator
- Combine only evidence-tagged findings while retaining uncertainty annotations
- Cap delegation depth and concurrent subagent count at the orchestration layer
- Include a deterministic aggregation policy for worker output acceptance
Explanation
Correct Answer
C:Limits on nesting and worker breadth bound cost and latency. Prompts alone are weaker than scheduler-level controls when an agent can keep delegating.
Incorrect Answers
A:Focused workers can receive targeted instructions and smaller context windows; this concerns matching specialist agents to distinct competencies, leaving the requirement for constraining recursive multi-agent delegation outstanding.
B:Outputs from specialists are proposals until verified; separating verified worker findings from guesses and constraining recursive multi-agent delegation are distinct responsibilities under these constraints.
D:A merge contract can specify completeness, conflict handling and evidence precedence; this can serve defining a reproducible worker merge policy, but leaves the requirement for constraining recursive multi-agent delegation unmet.
Question 200
A product owner describes an outcome but cannot decide between two incompatible data models. What should the agent do first?
- Examine references and present a removal plan before deletion
- Surface the architectural trade-off and request the missing business decision
- Obtain performance evidence and define the target metric before code changes
- Avoid moving from inspection to write-enabled execution
Explanation
Correct Answer
B:An unresolved design choice is not safe to infer from code alone. Planning should expose the decision to an authorized stakeholder before irreversible implementation.
Incorrect Answers
A:Deleting files on a hunch can break hidden consumers; that serves researching destructive refactors before execution, whereas the case requires resolving architecture ambiguity before editing.
C:Optimization without a baseline makes success impossible to judge and may harm unrelated paths; its benefits for baselining performance before optimization do not resolve the need for resolving architecture ambiguity before editing.
D:An assessment boundary prohibits side effects regardless of how confident the proposed fix appears; supporting honoring a read-only operating constraint cannot replace the required resolving architecture ambiguity before editing.