Question 30
A purchase approval agent occasionally chooses unnecessary tools although a fixed three-step sequence exists. Which next step is most defensible?
- Use a retrieval-augmented single-turn LLM design rather than autonomous agents.
- Implement a deterministic prompt-and-tool workflow with explicit stage transitions.
- Replace open-ended planning with a chained workflow for the known path.
- Use a bounded agent loop with explicit tool budgets and stop conditions.
Explanation
Correct Answer
C:Removing avoidable degrees of freedom makes the process easier to constrain, test and audit. Test whether the selected level of autonomy actually achieves preventing needless agent autonomy; extra agency is justified only by measurable gains in preventing needless agent autonomy.
Incorrect Answers
A:By itself, this improves bounded grounded question answering, but the case is not asking the architect to optimize that dimension. The requirement is preventing needless agent autonomy, supported by an approach to replace open-ended planning with a chained workflow for the known path.
B:Its main effect concerns predictable ordered processing. The case explicitly hinges on preventing needless agent autonomy; to resolve it reliably, the implementation has to replace open-ended planning with a chained workflow for the known path, which this choice does not establish.
D:The proposed action prioritizes adaptive research with bounded autonomy. That can coexist with the right solution, but it cannot replace preventing needless agent autonomy: the application must still replace open-ended planning with a chained workflow for the known path.
Question 31
A support assistant needs one database read plus a templated response under a strict median-latency target. Which approach should be recommended?
- Run independent validations in parallel and aggregate the verified results.
- Use a retrieval-augmented single-turn LLM design rather than autonomous agents.
- Use an evaluator-optimizer workflow with a measurable completion criterion.
- Prefer a single model step with a narrow retrieval call over a multi-agent team.
Explanation
Correct Answer
D:An orchestrator adds handoffs and context duplication that are not justified by this simple request. Test whether the selected level of autonomy actually achieves low-overhead design for simple tasks; extra agency is justified only by measurable gains in low-overhead design for simple tasks.
Incorrect Answers
A:This addresses parallel work on independent units; however, the problem asks for low-overhead design for simple tasks. Without the safeguard to prefer a single model step with a narrow retrieval call over a multi-agent team, the original failure mode is still possible.
B:It emphasizes bounded grounded question answering without closing the stated gap in low-overhead design for simple tasks. By contrast, the remedy to prefer a single model step with a narrow retrieval call over a multi-agent team acts on the condition that makes this scenario fail.
C:The proposed measure is aimed at bounded iterative quality improvement, not low-overhead design for simple tasks. The architecture must instead prefer a single model step with a narrow retrieval call over a multi-agent team; a separate change is still needed for low-overhead design for simple tasks.
Question 32
A fraud investigator's process branches unpredictably but certain monetary actions must remain outside model discretion. What is the best production design response?
- Combine agentic evidence collection with a deterministic transaction gate.
- Use a bounded agent loop with explicit tool budgets and stop conditions.
- Use an evaluator-optimizer workflow with a measurable completion criterion.
- Keep classification in a versioned deterministic decision engine; use Claude only for explanation.
Explanation
Correct Answer
A:Adaptive investigation can be separated from rule-governed actions so flexibility does not remove controls. Test whether the selected level of autonomy actually achieves hybrid agent-and-rules boundaries; extra agency is justified only by measurable gains in hybrid agent-and-rules boundaries.
Incorrect Answers
B:The relevant trade-off is not primarily adaptive research with bounded autonomy. A successful fix needs hybrid agent-and-rules boundaries; otherwise hybrid agent-and-rules boundaries remains unverified even after this change.
C:The control would be justified if bounded iterative quality improvement were the limiting factor. In this scenario the decisive target is hybrid agent-and-rules boundaries; success requires an architecture that will combine agentic evidence collection with a deterministic transaction gate.
D:The option belongs to the broader work on reproducible classification under fixed rules, but it cannot by itself deliver hybrid agent-and-rules boundaries. A decision on this case needs the team to combine agentic evidence collection with a deterministic transaction gate and verify that outcome.
Question 33
A customer-refund workflow first reserves store credit and then asks an external billing gateway to reverse a charge. Sometimes the reversal fails after the reservation, leaving the accounts inconsistent. Which workflow pattern best contains this partial-failure risk?
- Let the model choose a goodwill refund amount and hide the failed reversal from the record.
- Use explicit workflow states and compensating actions that release the reservation if the payment reversal fails.
- Restart the entire refund process after any error without recording completed side effects.
- Declare success once the assistant sends a confirmation message, regardless of payment-gateway state.
Explanation
Correct Answer
B:The operations span different systems and cannot be assumed to share one atomic commit. A saga-style state machine records which steps succeeded and invokes an authorized compensation for a later failure.
Incorrect Answers
A:Inventing a replacement financial decision does not restore consistency, and suppressing the failed action undermines the audit trail. Recovery should reverse or reconcile the actual committed step.
C:Blind restart can reserve credit twice and still fail to complete the reversal. Compensation requires a durable record of which business effects have already committed.
D:Text generation is not evidence of a committed financial operation. The workflow must use provider outcomes and verified compensating transitions before closing the case.
Question 34
A workflow invokes a second model to restate every answer but evaluation shows no quality gain. Which solution best resolves this operational gap?
- Eliminate the redundant model pass and retain targeted validation where risk warrants it.
- Replace open-ended planning with a chained workflow for the known path.
- Use an evaluator-optimizer workflow with a measurable completion criterion.
- Prefer a single model step with a narrow retrieval call over a multi-agent team.
Explanation
Correct Answer
A:Architecture complexity must earn its cost with measured quality or safety improvements. Test whether the selected level of autonomy actually achieves removing unjustified multi-step complexity; extra agency is justified only by measurable gains in removing unjustified multi-step complexity.
Incorrect Answers
B:This recommendation targets preventing needless agent autonomy at a different point in the workflow. It leaves removing unjustified multi-step complexity outside the design contract and does not require the team to eliminate the redundant model pass and retain targeted validation where risk warrants it.
C:This addresses bounded iterative quality improvement; however, the problem asks for removing unjustified multi-step complexity. Without the safeguard to eliminate the redundant model pass and retain targeted validation where risk warrants it, the original failure mode is still possible.
D:This would make sense as an additional project activity for low-overhead design for simple tasks. The priority required here is removing unjustified multi-step complexity, so the workflow still needs a control that will eliminate the redundant model pass and retain targeted validation where risk warrants it.
Question 35
An incident-analysis system has separate log, security and customer-impact investigations with independent evidence stores. What design decision would best fit these facts?
- Require explicit completion or timeout reconciliation before final synthesis.
- Use isolated agent workspaces and explicit structured handoff artifacts.
- Provision a scoped credential solely to that agent and pass a minimized result to the coordinator.
- Use an orchestrator to delegate scoped investigations to specialist agents and synthesize a reviewed conclusion.
- Log task assignments, tool calls, evidence references and synthesis decisions using shared trace identifiers.
Explanation
Correct Answer
D:Distinct tools and contexts justify task separation; central orchestration owns synthesis and final responsibility.
Incorrect Answers
A:The option belongs to the broader work on complete and timed aggregation, but it cannot by itself deliver bounded specialist delegation. A decision on this case needs the team to use an orchestrator to delegate scoped investigations to specialist agents and synthesize a reviewed conclusion and verify that outcome.
B:This shifts attention toward isolated state with accountable handoffs when the relevant decision is about bounded specialist delegation. Unless the architecture can use an orchestrator to delegate scoped investigations to specialist agents and synthesize a reviewed conclusion, the failure described in the stem remains unresolved.
C:For a workflow whose primary weakness was data minimization across specialist agents, that proposal could be reasonable. Here the unmet obligation is bounded specialist delegation, and that requires the system to use an orchestrator to delegate scoped investigations to specialist agents and synthesize a reviewed conclusion.
E:It is aimed at a different outcome, auditability of distributed reasoning. No evidence is given that it provides bounded specialist delegation; the needed mechanism is to use an orchestrator to delegate scoped investigations to specialist agents and synthesize a reviewed conclusion at the appropriate boundary.
Question 36
Several agents keep overwriting a shared investigation document and evidence provenance is lost. What should the solution architect prioritize?
- Use isolated agent workspaces and explicit structured handoff artifacts.
- Validate worker contracts at handoff and quarantine malformed worker output.
- Provision a scoped credential solely to that agent and pass a minimized result to the coordinator.
- Give the coordinator a conflict-resolution policy with evidence checks and escalation when unresolved.
Explanation
Correct Answer
A:Separation prevents state collisions while tagged evidence lets the coordinator reconcile conflicting findings.
Incorrect Answers
B:There is a legitimate use case for fault isolation at agent interfaces, but it does not supply the test or enforcement needed for isolated state with accountable handoffs. The design needs to use isolated agent workspaces and explicit structured handoff artifacts.
C:Consider its actual operating scope: data minimization across specialist agents. The requested assurance is isolated state with accountable handoffs, which calls for an implementation that can use isolated agent workspaces and explicit structured handoff artifacts, not merely for a related improvement.
D:The alternative could reduce risk associated with safe resolution of conflicting agent outputs, yet the described behavior requires isolated state with accountable handoffs. That outcome depends on the capability to use isolated agent workspaces and explicit structured handoff artifacts.
Question 37
A multi-agent deployment spends most of its time passing the same full transcript between workers. Which design choice best matches the stated constraint?
- Send only task-specific context and concise evidence references to each worker.
- Log task assignments, tool calls, evidence references and synthesis decisions using shared trace identifiers.
- Use isolated agent workspaces and explicit structured handoff artifacts.
- Bind tools and credentials to each worker's approved task scope.
Explanation
Correct Answer
A:Reducing irrelevant shared history avoids token duplication and cross-agent distraction. Assign a coordinator ownership of minimal agent handoff context and retain evidence and permission boundaries at each specialist handoff.
Incorrect Answers
B:By itself, this improves auditability of distributed reasoning, but the case is not asking the architect to optimize that dimension. The requirement is minimal agent handoff context, supported by an approach to send only task-specific context and concise evidence references to each worker.
C:Its main effect concerns isolated state with accountable handoffs. The case explicitly hinges on minimal agent handoff context; to resolve it reliably, the implementation has to send only task-specific context and concise evidence references to each worker, which this choice does not establish.
D:The option might improve per-agent permission boundaries; that remains insufficient for minimal agent handoff context. The case is resolved by minimal agent handoff context, which depends on a design that will send only task-specific context and concise evidence references to each worker.
Question 38
Two specialists disagree about whether a file contains an export-control violation. Which architecture adjustment is most appropriate?
- Send only task-specific context and concise evidence references to each worker.
- Log task assignments, tool calls, evidence references and synthesis decisions using shared trace identifiers.
- Use isolated agent workspaces and explicit structured handoff artifacts.
- Give the coordinator a conflict-resolution policy with evidence checks and escalation when unresolved.
Explanation
Correct Answer
D:Agent voting alone is not a compliance decision; conflicting claims need adjudication against authoritative evidence.
Incorrect Answers
A:Although minimal agent handoff context can be valuable, it is not the acceptance condition in the question. The missing provision is safe resolution of conflicting agent outputs, addressed by the control to give the coordinator a conflict-resolution policy with evidence checks and escalation when unresolved.
B:This addresses auditability of distributed reasoning; however, the problem asks for safe resolution of conflicting agent outputs. Without the safeguard to give the coordinator a conflict-resolution policy with evidence checks and escalation when unresolved, the original failure mode is still possible.
C:It emphasizes isolated state with accountable handoffs without closing the stated gap in safe resolution of conflicting agent outputs. By contrast, the remedy to give the coordinator a conflict-resolution policy with evidence checks and escalation when unresolved acts on the condition that makes this scenario fail.
Question 39
A financial agent delegates to another worker that can initiate transfers despite having only a research mandate. Which next step is most defensible?
- Cap delegation depth, child count and per-job resource budgets.
- Log task assignments, tool calls, evidence references and synthesis decisions using shared trace identifiers.
- Bind tools and credentials to each worker's approved task scope.
- Send only task-specific context and concise evidence references to each worker.
Explanation
Correct Answer
C:Least-privilege tool exposure stops delegated workers from inheriting an orchestrator's broad authority. Assign a coordinator ownership of per-agent permission boundaries and retain evidence and permission boundaries at each specialist handoff.
Incorrect Answers
A:The alternative could reduce risk associated with controlled multi-agent resource use, yet the described behavior requires per-agent permission boundaries. That outcome depends on the capability to bind tools and credentials to each worker's approved task scope.
B:The proposed measure is aimed at auditability of distributed reasoning, not per-agent permission boundaries. The architecture must instead bind tools and credentials to each worker's approved task scope; a separate change is still needed for per-agent permission boundaries.
D:The control would be justified if minimal agent handoff context were the limiting factor. In this scenario the decisive target is per-agent permission boundaries; success requires an architecture that will bind tools and credentials to each worker's approved task scope.
Question 40
A research orchestrator keeps creating new subagents for minor subtasks until its token budget is exhausted. Which approach should be recommended?
- Validate worker contracts at handoff and quarantine malformed worker output.
- Send only task-specific context and concise evidence references to each worker.
- Cap delegation depth, child count and per-job resource budgets.
- Use isolated agent workspaces and explicit structured handoff artifacts.
Explanation
Correct Answer
C:Structured limits protect cost and latency without assuming the model will spontaneously stop exploring. Assign a coordinator ownership of controlled multi-agent resource use and retain evidence and permission boundaries at each specialist handoff.
Incorrect Answers
A:The option might improve fault isolation at agent interfaces; that remains insufficient for controlled multi-agent resource use. The case is resolved by controlled multi-agent resource use, which depends on a design that will cap delegation depth, child count and per-job resource budgets.
B:The proposed action prioritizes minimal agent handoff context. That can coexist with the right solution, but it cannot replace controlled multi-agent resource use: the application must still cap delegation depth, child count and per-job resource budgets.
D:This addresses isolated state with accountable handoffs; however, the problem asks for controlled multi-agent resource use. Without the safeguard to cap delegation depth, child count and per-job resource budgets, the original failure mode is still possible.