Question 1
An invoice process must extract fields, run a fixed tax check, and request approval in that order. Every invoice follows the same stages. Which architecture provides the clearest control over this sequence?
- Let Claude choose the stage order separately for each invoice.
- Chain extraction, tax calculation, and approval prompts, allowing each output to advance without a programmatic check.
- Use an explicit workflow with validation gates between stages.
- Ask several independent agents to complete the whole process and select a majority result.
Explanation
Correct Answer
C:Code can enforce the required transitions and stop an invalid invoice before approval. Claude can still perform extraction within a stage without deciding whether mandatory stages occur.
Incorrect Answers
A:Dynamic planning introduces variability where the process already defines the dependencies. A model might skip or reorder a required check unless the application constrains execution.
B:Prompt chaining follows a sequence, but an ungated chain can forward incomplete or invalid intermediate output. The requested control includes verifying that each prerequisite succeeded before advancing.
D:Voting compares outputs but does not establish that every required control ran in order. Agreement on an invoice does not demonstrate compliance with the processing sequence.
Question 2
A refactoring coordinator discovers affected modules only after examining the dependency graph. Some modules need specialists; others need no changes. Which orchestration design accommodates this variable scope while avoiding unnecessary workers?
- Run a fixed chain containing exactly one worker for each of three predefined modules.
- Route the request to one language specialist using only the issue title.
- Create a worker for every repository directory before inspecting the graph.
- Let the coordinator inspect dependencies, define the required subtasks, and delegate them with explicit return contracts.
Explanation
Correct Answer
D:The orchestrator derives the work breakdown from the actual dependency evidence. Specialists handle the discovered subtasks, while the coordinator remains responsible for integration and deciding whether coverage is complete.
Incorrect Answers
A:A fixed decomposition is effective when the required subtasks are known. Here the affected modules emerge from inspection, so the chain can omit necessary work or schedule irrelevant work.
B:A language classifier can select expertise, but it cannot discover the changing set of cross-module tasks. The issue title lacks the dependency evidence needed to define that work.
C:This covers potential modules but commits resources before determining relevance. It does not meet the requirement to avoid unnecessary workers for unaffected directories.
Question 3
A support system has separate billing, account-recovery, and integration procedures. Each request belongs to one category, and each procedure needs different tools. How should the architect avoid presenting every specialist workflow to every request?
- Provide every procedure and tool in a shared prompt, then let the selected specialist ignore the rest.
- Classify the request and dispatch it to the relevant specialist workflow.
- Select the specialist with the shortest queue without considering request category.
- Run each request through all three specialist procedures in sequence.
Explanation
Correct Answer
B:Routing uses the request category to select the appropriate handling path. Each specialist can receive a focused prompt and tool set, reducing interference from unrelated procedures.
Incorrect Answers
A:This can sometimes produce a useful answer, but it still presents every workflow to every request. It fails the explicit isolation goal and increases the burden on tool selection.
C:Load balancing optimizes utilization among suitable handlers. Choosing an unsuitable specialist by queue length does not supply the procedure or tools needed for that request.
D:Serial composition can help when every stage is required. These requests belong to mutually exclusive categories, so the sequence introduces irrelevant procedures and tool exposure.
Question 4
A report needs independent reviews of three regional datasets. A final narrative must compare all three reviews. Which execution plan reduces avoidable waiting without violating dependencies?
- Start the comparison with the first completed review and publish without another synthesis pass.
- Run the regional reviews concurrently, then synthesize after collecting their results.
- Wait for two of the three reviews, then publish the majority conclusion.
- Run each regional review only after the previous review completes.
Explanation
Correct Answer
B:Independent analyses can overlap in time. A join before synthesis ensures the final comparison has the evidence from every region rather than relying on incomplete intermediate results.
Incorrect Answers
A:Early drafting can reduce latency if a later pass incorporates all evidence. Publishing without that pass leaves later regional findings outside the comparison.
C:A quorum can serve some voting tasks, but the report must compare every region. Two independent datasets cannot stand in for the missing third region.
D:Serial execution respects correctness but introduces unnecessary waiting because no review consumes another region's output. The scenario explicitly permits independent work.
Question 5
A coordinator asks a fresh, non-forked research worker to investigate a failed migration. The worker cannot see the coordinator's conversation. What should the delegation message contain?
- Send the coordinator's suspected cause and request only evidence supporting it.
- Send the failure label and tell the worker to use constraints mentioned earlier in the parent conversation.
- The failure evidence, relevant constraints, source locations, and expected return format.
- Send source locations and ask for a general migration summary without specifying the observed failure.
Explanation
Correct Answer
C:A self-contained task packet supplies the information the worker needs to investigate independently. An explicit return contract also makes the result easier for the coordinator to evaluate and combine.
Incorrect Answers
A:This creates a confirmation-biased assignment. A useful investigation must test the hypothesis against contrary observations as well as supporting evidence.
B:The parent history is unavailable to this worker. Referring to earlier constraints without providing them leaves the assignment dependent on inaccessible information.
D:The worker can inspect sources, but a general summary need not investigate the actual incident. Failure evidence and a focused outcome distinguish diagnosis from background research.
Question 6
A main agent repeatedly loses track of its task after reading thousands of diagnostic log lines. Only a short root-cause summary is needed for the final decision. Which change directly limits this context pollution?
- Enable prompt caching for the raw logs while continuing to include all of them in the main conversation.
- Have a dedicated worker inspect the logs and return relevant findings with evidence references.
- Have the main agent ingest every log, then summarize them before writing the final answer.
- Assign the logs to a worker that returns the entire raw log stream to the main agent.
Explanation
Correct Answer
B:The bulky inspection can occur outside the main conversation, while a concise result preserves the information needed for the decision. Evidence references allow follow-up without importing the entire log stream.
Incorrect Answers
A:Caching can reduce repeated input processing. It does not remove the logs from the model's working context or identify the root-cause evidence among them.
C:A later summary may shorten retained history, but the main conversation must first absorb the entire diagnostic stream. It does not isolate the inspection that is causing the stated problem.
D:Delegation isolates the worker's intermediate activity, but returning all raw material recreates the context burden. The return contract should focus on findings and evidence locations.
Question 7
A refund service verifies that an approval record exists. An agent then reuses approval for order X to request a larger refund on order Y. Which missing check directly prevents this misuse?
- Require the agent to include a description of the approval in the request.
- Verify only that the approval record was created recently.
- Check that the approver belongs to the refunds team, without comparing transaction details.
- Match the approved order, amount, currency, and permitted use against the requested transaction at commit time.
Explanation
Correct Answer
D:The service must verify the scope of the actual approval, including any replay restrictions. Binding it to the transaction prevents a valid record from being repurposed for a different refund.
Incorrect Answers
A:A description gives an audit trail of what the agent claims. It does not independently bind the authorized transaction to the requested refund.
B:Freshness limits the age of an approval but does not establish what it authorizes. A recent approval for order X still cannot justify a different order or amount.
C:An authorized approver is necessary but insufficient. Their authority does not mean every request made after one approval falls within that approval's scope.
Question 8
A create-ticket request times out after submission. The service guarantees that retries with the same idempotency key and payload return the original operation within its retention window. The retry is within that window. What recovery avoids duplicate tickets?
- Retry the same payload with a fresh key after a short delay.
- Reuse the original idempotency key and reconcile the operation's recorded status.
- Search for a matching ticket title and create another ticket if no match appears immediately.
- Retry without a key and ask the service to deduplicate by description.
Explanation
Correct Answer
B:A stable operation identity lets the service recognize a retry as the same intended creation. Status reconciliation distinguishes a completed write from one that still needs execution without blindly duplicating it.
Incorrect Answers
A:Waiting may help a transient outage, but a fresh key identifies a new operation. If the original request committed, the service can create another ticket.
C:Search indexes can lag, and different tickets can share titles. The absence of an immediate match is weaker evidence than the service's guaranteed operation identity.
D:The stated guarantee depends on the original key and payload. Content-based deduplication has not been specified and cannot be assumed to provide equivalent behavior.
Question 9
An investigator keeps calling search with slightly different wording but finds no new evidence. Which termination policy best prevents an expensive loop while preserving a useful outcome?
- Stop whenever the model reports high confidence, without checking whether the evidence changed.
- Stop only when a query returns exactly the same URL list as its immediate predecessor.
- Reset the iteration counter whenever the search wording changes.
- Apply iteration and cost budgets, detect lack of progress, and return unresolved findings when limits are reached.
Explanation
Correct Answer
D:External bounds make termination predictable, while progress checks identify repeated work earlier. A partial finding with explicit uncertainty is more trustworthy than inventing a conclusion to satisfy a completion signal.
Incorrect Answers
A:Confidence is not an external measure of progress. Repeated low-value observations can still lead to a confident answer, so this policy does not reliably bound unproductive search.
B:A loop can alternate between queries or reorder equivalent results while yielding no new evidence. Adjacent-list equality misses broader repetition and supplies no hard cost limit.
C:Cosmetic query changes would continually renew the budget. Progress should reflect additional relevant evidence or completed work, rather than merely different input text.
Question 10
A customer asks to change the destination of an active shipment. The carrier response states that the package has entered a restricted handoff stage and the available tool cannot change it. What should the assistant do next?
- Close the request as impossible without passing it to the available human exception process.
- Update the customer's saved address and treat that as redirecting the active package.
- Explain the tool limitation and escalate through the supported human process with the relevant shipment evidence.
- Retry the same change using exponential backoff for the rest of the handoff period.
Explanation
Correct Answer
C:The requested change exceeds the available execution path. Escalation preserves the observed carrier state and allows an authorized operator to determine whether an exceptional remedy exists.
Incorrect Answers
A:The tool limitation establishes that this automated path cannot act. It does not establish that the supported human process has no remedy, so closing the request prematurely discards that path.
B:A profile update can affect future shipments without changing the carrier's active routing record. The restricted package still needs the carrier's supported exception process.
D:Backoff addresses temporary availability problems. The response describes a business-state restriction, so repeated calls do not supply the unavailable redirection capability.