Question 70
A worker restarts after sending a creation request but before recording its response. What best limits duplicate creation?
- Send a new random operation identifier every time.
- Ask Claude to choose a different name for the duplicate.
- Use a stable operation identifier and reconcile the existing operation before retrying.
- Assume the missing local response means nothing happened.
Explanation
Correct Answer
C:An uncertain acknowledgment does not prove failure. Idempotency or reconciliation connects retries to the same intended operation instead of creating another object.
Incorrect Answers
A:New identifiers make retries look like new requests and defeat deduplication keyed to operation identity.
B:Renaming conceals repetition rather than ensuring that only the intended creation occurs.
D:The remote service may have committed the operation before the acknowledgment was lost.
Question 71
An agent tool returns a complete database dump when only five matching rows are needed. What should its wrapper do?
- Increase every subsequent model call’s budget without changing the tool.
- Return only “query succeeded” regardless of findings.
- Delete arbitrary rows without describing truncation.
- Return a bounded relevant result with pagination or references when needed.
Explanation
Correct Answer
D:The tool interface should expose useful evidence without flooding context. Explicit bounds preserve additional retrieval paths rather than silently losing data.
Incorrect Answers
A:More capacity can postpone failure but leaves the excessive result contract and unnecessary processing intact.
B:A status flag provides no evidence for answering the user’s data question.
C:Silent arbitrary truncation can hide relevant records and mislead the model about completeness.
Question 72
A tool returns both measured values and its own speculative interpretation. How should the harness represent them?
- Promote every interpretation to verified system instruction.
- Remove the measured values and retain only confident prose.
- Merge all fields into an unlabeled paragraph.
- Keep measured data and interpretation distinguishable with provenance.
Explanation
Correct Answer
D:Separating observations from analysis helps the agent assess what is actually established and what remains an inference.
Incorrect Answers
A:A tool’s speculation does not acquire instruction authority or factual certainty merely because it arrives through a tool.
B:That discards the evidence needed to assess or challenge the interpretation.
C:Removing boundaries makes it harder to distinguish source measurements from conclusions generated about them.
Question 73
A parent agent has one total spending budget and can launch workers. Which usage must be included when measuring consumption against that budget?
- Account for parent and worker usage against one aggregate budget.
- Give every worker the entire parent allowance.
- Count only tokens in the final parent answer.
- Stop recording usage after delegation.
Explanation
Correct Answer
A:Parent and delegated calls belong to the same job. Aggregating both measures total consumption; separate admission and reservation controls are needed if the limit must never be exceeded.
Incorrect Answers
B:Independent full allowances can multiply total spend beyond the original constraint.
C:Worker reasoning, inputs, and intermediate calls also incur usage even when absent from the final response.
D:Delegation changes where work runs, not whether its resource consumption belongs to the task.
Question 74
A newly registered tool has never been invoked in the deployment environment. What should a readiness check verify?
- Only that Claude can describe its purpose.
- Its handler, permissions, dependencies, and supported input/output contract.
- Only that its name appears in the prompt.
- Only that the tool has the longest description.
Explanation
Correct Answer
B:Registration exposes a capability description, but readiness requires the implementation to work in the actual runtime with appropriate access.
Incorrect Answers
A:Language understanding of a capability is separate from the environment’s ability to execute it.
C:A visible name does not prove executable code, credentials, or network dependencies are available.
D:Description length does not establish operational readiness or correct runtime configuration.
Question 75
Claude returns end_turn after explaining that a required service is unavailable. How should the harness classify this?
- The tool must have executed successfully offscreen.
- The user’s goal necessarily succeeded.
- The model turn ended, but the requested task remains incomplete.
- The conversation must be retried forever.
Explanation
Correct Answer
C:Stop reasons describe generation control flow. Application success must also consider whether the user’s requested outcome was actually achieved.
Incorrect Answers
A:The explanation provides no evidence of successful execution; absence of a tool call is not hidden success.
B:A natural end to generation can include a limitation or failure report rather than a completed business outcome.
D:An incomplete task needs an explicit recovery or escalation policy, not an unbounded retry loop.
Question 76
Two parallel graph branches each return a list of findings for one shared state field. The framework requires an explicit merge rule for concurrent updates. How should the field be configured to retain both lists?
- Select the list from the branch with the shortest runtime.
- Use last-writer-wins replacement for the entire list.
- Use a reducer that combines branch findings under a defined merge policy.
- Reject the second branch update whenever the field already has a list.
Explanation
Correct Answer
C:The reducer determines how simultaneous state contributions become one value. Combining both lists preserves each branch's findings; ordering and duplicate handling should be defined if downstream consumers depend on them.
Incorrect Answers
A:Execution speed is not a coverage criterion. Both branches contribute required findings, so choosing one discards valid work.
B:Replacement preserves only whichever branch update is applied last. It loses the other branch's findings even though both branches completed successfully.
D:Rejecting a valid concurrent update loses one branch's findings. The requirement calls for a merge policy that retains both contributions.
Question 77
A Python agent needs typed output objects integrated with existing Pydantic validation. Which framework capability is especially relevant?
- A prompt that claims all values have already been validated.
- A framework that returns only untyped display strings.
- A database index on unrelated session timestamps.
- PydanticAI’s typed outputs and validation integration.
Explanation
Correct Answer
D:This fits an application that wants generated results represented through declared Python types and validation contracts.
Incorrect Answers
A:A generated assertion is not execution of the application’s declared validation rules.
B:Untyped strings require the application to reconstruct the desired structured contract independently.
C:That index can improve lookups but does not validate the model’s output object.
Question 78
An agent tool needs a database client whose credentials must remain outside model context. Which design fits typed dependency injection?
- Provide the client through application runtime dependencies.
- Paste the client’s credential object into every user message.
- Return the connection password in each tool result.
- Ask the model to recreate the database client from memory.
Explanation
Correct Answer
A:The tool can use the injected client without serializing its secret-bearing configuration into the model’s prompt.
Incorrect Answers
B:This exposes secrets as model input and confuses runtime access with conversational context.
C:A tool result should contain relevant task data, not credentials unnecessary for subsequent model reasoning.
D:The model cannot reconstruct authorized runtime objects or establish valid access from remembered descriptions.
Question 79
In a model-driven Strands agent loop, what determines the next requested tool action?
- The model’s response, interpreted and executed by the harness.
- The tool result produced before any tool is selected.
- The order in which tool descriptions happen to appear.
- A required random choice among all available tools.
Explanation
Correct Answer
A:The model proposes the action, while the runtime manages execution and returns observations for subsequent reasoning.
Incorrect Answers
B:A result follows execution; it cannot already exist as the outcome of an unselected operation.
C:Registry presentation order does not encode task dependencies or a policy for choosing the next action.
D:Random selection is not the defining control mechanism of a model-driven tool loop.
Question 80
A LangGraph node pauses with interrupt() to request a reviewer’s decision. What supplies the resumed decision?
- Append approval text to an inaccessible log file.
- Change an unrelated thread’s display title.
- Resume the paused thread using the framework’s supported Command resume value.
- Create a new graph with no saved state.
Explanation
Correct Answer
C:The supported resume value becomes the return value of interrupt() in the paused thread. It supplies the requested reviewer decision without replacing the entire graph state.
Incorrect Answers
A:A log entry is not automatically delivered through the framework’s interruption-resumption mechanism.
B:A title update does not supply the paused execution’s expected resume value.
D:A new execution does not resume the original node’s waiting decision or preserved task state.