Question 50
A workflow needs exactly-once-looking external business operations though retries are unavoidable. What combination is most practical?
- Use durable state and idempotent tool operations across retries.
- Persist the workflow position and completed side effects outside process memory.
- Retain a compact, explicit summary of durable decisions and retrieve details when needed.
- For this workflow, demonstrate a needed capability such as durable state, branching, or observability.
Explanation
Correct Answer
A:Recovery requires both knowledge of committed stages and protection from duplicate writes; neither a model summary nor blind retries supplies that guarantee.
Incorrect Answers
B:Durable checkpoints let the executor resume without blindly rerunning earlier writes after a crash; this targets crash-safe workflow checkpoints, while the system must demonstrate side-effect safe recovery.
C:Separating long-term task facts from raw transcript prevents context waste while preserving the information needed for new decisions; the effect concerns compact durable memory; the success condition is side-effect safe recovery.
D:Framework overhead is worthwhile when its orchestration features solve a real problem, not merely because agents are fashionable; even if framework selection by need succeeds, side-effect safe recovery still needs separate verification.
Question 51
A manager wants three workers to edit the same file concurrently. What scheduling constraint should it add?
- Partition ownership or serialize writes to shared artifacts.
- At the application layer, compute aggregates with a tool and pass only validated results back to Claude.
- Persist stable configuration in versioned files and load it into each session.
- As an alternative, route bounded subtasks to domain-specific workers with narrow toolsets.
- Check that the framework exposes enforceable permission or pre-tool hooks.
Explanation
Correct Answer
A:Parallel reasoning can be safe, but simultaneous edits to common state risk conflicts and lost changes unless ownership is coordinated.
Incorrect Answers
B:Moving mechanical data processing into code avoids token-heavy raw inputs while keeping Claude focused on interpretation; this implements offloading deterministic computation rather than the demanded shared-artifact concurrency.
C:Reproducible configuration should be external to ephemeral conversation memory, making changes reviewable and repeatable; the shared-artifact concurrency criterion differs from versioned durable configuration; the gap persists.
D:Specialized contexts and tools lower irrelevant information and permission exposure, while a parent assembles the answer; the scenario's controlling requirement is shared-artifact concurrency, not the related domain-based subagent routing concern.
E:Framework convenience cannot substitute for mandatory execution controls; approval must intercept the actual write operation; testing shared-artifact concurrency would still fail after improving framework approval integration.
Question 52
A repository agent needs configuration that is consistent across restarts, not something inferred from chat. Where should authoritative state live?
- Grant the web-reading worker limited tools and treat its findings as data, not instructions.
- Store stable configuration in versioned files and load it into each session.
- Review the execution trace and resulting system state against task criteria.
- Monitor repeated action-observation signatures and terminate or replan on repetition.
Explanation
Correct Answer
B:Reproducible configuration should be external to ephemeral conversation memory, making changes reviewable and repeatable.
Incorrect Answers
A:Isolating untrusted retrieval prevents document text from inheriting authority over high-impact actions in the parent process; the effect concerns untrusted research containment; the success condition is versioned durable configuration.
C:Multi-step success depends on tool choices and side effects, not just fluent output; trajectory evidence reveals hidden failures; the scenario's controlling requirement is versioned durable configuration, not the related state-based agent evaluation concern.
D:Cycle detection identifies lack of progress independently of the model's confidence and prevents unbounded redundant calls; for versioned durable configuration, this instead manages repeated action loop detection; the requested decision remains unresolved.
Question 53
An agent's task spans three separate user requests over a week. How should it remember progress?
- Divide ownership or serialize writes to shared artifacts.
- Preserve a compact, explicit summary of durable decisions and retrieve details when needed.
- Deliver a constrained result object containing findings, confidence, and evidence IDs.
- Persist task status and identifiers in an external store rather than replaying every token.
Explanation
Correct Answer
D:Durable state records progress across sessions and can be validated; full transcript replay is costly and may not reflect authoritative system state.
Incorrect Answers
A:Parallel reasoning can be safe, but simultaneous edits to common state risk conflicts and lost changes unless ownership is coordinated; for cross-session task continuity, this instead manages shared-artifact concurrency; the requested decision remains unresolved.
B:Separating long-term task facts from raw transcript prevents context waste while preserving the information needed for new decisions; where cross-session task continuity matters, a control for compact durable memory leaves that test unmet.
C:A structured contract limits what crosses the context boundary and provides information the coordinator can verify; changing structured specialist handoff does not establish cross-session task continuity in this application.
Question 54
A planner needs a specialist to inspect untrusted web pages. What boundary should the manager apply?
- Maintain source identifiers and selected excerpts in structured working memory.
- Keep stable configuration in versioned files and load it into each session.
- Compact history while explicitly retaining only still-valid instructions and state.
- Give the web-reading worker limited tools and treat its findings as data, not instructions.
Explanation
Correct Answer
D:Isolating untrusted retrieval prevents document text from inheriting authority over high-impact actions in the parent process.
Incorrect Answers
A:Evidence provenance makes later claims checkable; keeping only an unreferenced summary would lose the link to the supporting records; because provenance in agent memory is its concern, untrusted research containment remains outside its effect here.
B:Reproducible configuration should be external to ephemeral conversation memory, making changes reviewable and repeatable; for untrusted research containment, a control centered on versioned durable configuration is not the appropriate substitute.
C:Selective compaction prevents stale turns from silently competing with current requirements; blindly appending history preserves drift; the scenario's controlling requirement is untrusted research containment, not the related multi-criterion integration concern.
Question 55
A team measures agent quality solely by whether a final paragraph sounds convincing. What should change?
- Validate that the framework exposes enforceable permission or pre-tool hooks.
- Instruct the worker for a bounded evidence summary with references to the originals.
- Evaluate the execution trace and resulting system state against task criteria.
- Retain source identifiers and selected excerpts in structured working memory.
Explanation
Correct Answer
C:Multi-step success depends on tool choices and side effects, not just fluent output; trajectory evidence reveals hidden failures.
Incorrect Answers
A:Framework convenience cannot substitute for mandatory execution controls; approval must intercept the actual write operation; for state-based agent evaluation, a control centered on framework approval integration is not the appropriate substitute.
B:Compact outputs preserve relevant facts and sources while preventing the parent context from being overwhelmed by intermediate material; when the goal is state-based agent evaluation, optimizing concise worker return contracts misses that operational boundary.
D:Evidence provenance makes later claims checkable; keeping only an unreferenced summary would lose the link to the supporting records; even if provenance in agent memory succeeds, state-based agent evaluation still needs separate verification.
Question 56
An assistant repeatedly passes a 10,000-line CSV through the conversation although only aggregates matter. What design is more efficient?
- Record the workflow position and completed side effects outside process memory.
- Summarize history while explicitly retaining only still-valid instructions and state.
- Distinguish session scratch state from durable, reviewed memory records.
- Compute aggregates with a tool and pass only validated results back to Claude.
Explanation
Correct Answer
D:Moving mechanical data processing into code avoids token-heavy raw inputs while keeping Claude focused on interpretation.
Incorrect Answers
A:Durable checkpoints let the executor resume without blindly rerunning earlier writes after a crash; the effect concerns crash-safe workflow checkpoints; the success condition is offloading deterministic computation.
B:Selective compaction prevents stale turns from silently competing with current requirements; blindly appending history preserves drift; the consequence is improved multi-criterion integration, not verified offloading deterministic computation.
C:Temporary observations should expire with the task; promoting only durable facts reduces stale or sensitive data accumulation; changing transient versus durable memory does not establish offloading deterministic computation in this application.
Question 57
A specialized worker must not disclose internal chain-of-work to the parent, only validated facts. What interface fits?
- For this workflow, route bounded subtasks to domain-specific workers with narrow toolsets.
- Carry out the critic in a separate context with the draft and evaluation rubric.
- Record repeated action-observation signatures and terminate or replan on repetition.
- Return a constrained result object containing findings, confidence, and evidence IDs.
- Assign the web-reading worker limited tools and treat its findings as data, not instructions.
Explanation
Correct Answer
D:A structured contract limits what crosses the context boundary and provides information the coordinator can verify.
Incorrect Answers
A:Specialized contexts and tools lower irrelevant information and permission exposure, while a parent assembles the answer; where structured specialist handoff matters, a control for domain-based subagent routing leaves that test unmet.
B:Isolated critique reduces anchoring to exploratory reasoning and yields a more independent check of the artifact; it follows independent evaluator context constraints but leaves structured specialist handoff unspecified.
C:Cycle detection identifies lack of progress independently of the model's confidence and prevents unbounded redundant calls; a remedy for repeated action loop detection is not a substitute for structured specialist handoff.
E:Isolating untrusted retrieval prevents document text from inheriting authority over high-impact actions in the parent process; the structured specialist handoff criterion differs from untrusted research containment; the gap persists.
Question 58
A service must distinguish temporary working notes from information that should persist indefinitely. Which memory approach is appropriate?
- Adopt durable state and idempotent tool operations across retries.
- Log repeated action-observation signatures and terminate or replan on repetition.
- Maintain task status and identifiers in an external store rather than replaying every token.
- Separate session scratch state from durable, reviewed memory records.
Explanation
Correct Answer
D:Temporary observations should expire with the task; promoting only durable facts reduces stale or sensitive data accumulation.
Incorrect Answers
A:Recovery requires both knowledge of committed stages and protection from duplicate writes; neither a model summary nor blind retries supplies that guarantee; for this request, transient versus durable memory is the missing piece, not side-effect safe recovery.
B:Cycle detection identifies lack of progress independently of the model's confidence and prevents unbounded redundant calls; when the goal is transient versus durable memory, optimizing repeated action loop detection misses that operational boundary.
C:Durable state records progress across sessions and can be validated; full transcript replay is costly and may not reflect authoritative system state; this may help cross-session task continuity but cannot by itself deliver transient versus durable memory.
Question 59
A framework generates its own tool loop, but the organization mandates approval before any write. What should determine whether it can be adopted?
- Verify that the framework exposes enforceable permission or pre-tool hooks.
- Have the worker for a bounded evidence summary with references to the originals.
- Adopt a workflow graph with explicit states and transition rules.
- Record task status and identifiers in an external store rather than replaying every token.
Explanation
Correct Answer
A:Framework convenience cannot substitute for mandatory execution controls; approval must intercept the actual write operation.
Incorrect Answers
B:Compact outputs preserve relevant facts and sources while preventing the parent context from being overwhelmed by intermediate material; the effect concerns concise worker return contracts; the success condition is framework approval integration.
C:Graph-based orchestration makes conditional edges and retry points inspectable, rather than hiding all sequencing in natural-language planning; this targets inspectable task-state graph, while the system must demonstrate framework approval integration.
D:Durable state records progress across sessions and can be validated; full transcript replay is costly and may not reflect authoritative system state; for framework approval integration, this instead manages cross-session task continuity; the requested decision remains unresolved.
Question 60
A helpdesk sponsor asks for an AI assistant that is 'accurate and fast.' What should engineering define before choosing a model?
- Specify supported languages and measure task quality for each representative locale.
- Approximate document volume, query rate, indexing load, and latency budgets.
- Convert the request into measurable quality, latency, throughput, and risk targets.
- Evaluate source-supported claims and rejection of unsupported questions on a held-out set.
Explanation
Correct Answer
C:Acceptance criteria turn adjectives into testable constraints; model selection without task metrics can optimize the wrong property.
Incorrect Answers
A:Language performance varies by input distribution and task type, so a single English evaluation cannot establish fitness; it follows multilingual quality targets constraints but leaves measurable product requirements unspecified.
B:Retrieval storage and preprocessing costs differ from model inference; capacity planning needs both workloads; the effect concerns retrieval infrastructure sizing; the success condition is measurable product requirements.
D:Grounding is a product behavior that needs verification against retrieval evidence, not a vague instruction to avoid hallucinations; this may help grounded answer acceptance but cannot by itself deliver measurable product requirements.