Question 50
A task contains thirty trivial string replacements, each cheaper to perform than to coordinate through another model call. Which approach is most efficient?
- Choose the largest model for each individual replacement.
- Create one specialist agent for every replacement.
- Ask each specialist to rediscover the replacement rule.
- Execute the replacements in ordinary code within one task.
Explanation
Correct Answer
D:Deterministic transformations need no separate reasoning agent. Local code avoids delegation overhead while producing directly testable edits.
Incorrect Answers
A:Model size does not justify repeated inference for a simple deterministic operation. The specified transformation can be implemented directly.
B:Coordination dominates the useful work in this scenario. Dividing the task further increases cost without adding meaningful reasoning capability.
C:The rule is already known. Re-deriving it independently wastes effort and creates opportunities for inconsistent interpretation.
Question 51
Two coding agents must edit overlapping lines in one file. Which coordination choice most directly avoids lost updates?
- Tell each agent that it owns the whole file exclusively.
- Serialize overlapping edits or merge isolated changes with conflict checks.
- Allow both to overwrite the file and keep the last write.
- Increase the model context limit for both agents.
Explanation
Correct Answer
B:The shared write target creates a real conflict risk. Controlled sequencing or explicit merging prevents one agent’s output from silently replacing another’s work.
Incorrect Answers
A:Contradictory ownership claims do not establish actual exclusion. Both processes may still modify the same content concurrently.
C:The last writer may discard valid earlier changes. Completion order is not a substitute for checking whether edits are compatible.
D:A larger context does not provide filesystem locking or merge resolution. The failure is write coordination rather than input capacity.
Question 52
Independent checks take approximately two, five, and eight seconds. Ignore coordination overhead. What determines completion time when all three run concurrently and all are required?
- Approximately the longest check, eight seconds.
- Approximately their sum, fifteen seconds.
- Approximately the shortest check, two seconds.
- Exactly zero seconds after scheduling.
Explanation
Correct Answer
A:The combined task waits for every required result. With simultaneous independent execution, the slowest check determines the critical path under the stated assumptions.
Incorrect Answers
B:Summing durations describes sequential execution. It does not reflect the overlap explicitly allowed by concurrent independent checks.
C:The fastest result is insufficient when every check is required. The remaining checks still need to complete before the combined review finishes.
D:Scheduling starts work; it does not produce completed results. The application must still await the required operations.
Question 53
A troubleshooting task may require different diagnostic tools depending on each new observation. The tool sequence cannot be enumerated beforehand. Which approach is suitable?
- A bounded agent that chooses subsequent diagnostics from observations.
- An unbounded loop with unrestricted administrative tools.
- A fixed sequence that ignores diagnostic results.
- A single static answer generated before any diagnostics.
Explanation
Correct Answer
A:The task requires adapting the next step to new evidence. A controlled agent loop supports that flexibility while retaining budgets and explicit execution boundaries.
Incorrect Answers
B:Adaptability does not require unlimited authority or execution. Removing bounds introduces risk without improving the evidential basis for choosing diagnostics.
C:Ignoring observations removes the adaptation the task requires. The sequence may perform irrelevant checks while missing newly indicated causes.
D:The answer would lack the evidence needed to distinguish possible failures. The scenario requires investigation rather than an unsupported initial conclusion.
Question 54
An assistant must total validated invoice amounts exactly before explaining the result. Which component should perform the arithmetic?
- A more persuasive prompt asking for perfect arithmetic.
- Several language-model guesses followed by majority voting.
- The response formatter by concatenating amount strings.
- A deterministic calculation routine using the validated numeric values.
Explanation
Correct Answer
D:Code can apply the specified arithmetic precisely and expose a testable result. Claude can then explain the calculation without being the sole arithmetic authority.
Incorrect Answers
A:The wording expresses a goal but provides no deterministic guarantee. Exact computation should use an appropriate calculation mechanism.
B:Consensus does not guarantee exact arithmetic. Repeated guesses add inference cost instead of using the deterministic computation already available.
C:String concatenation joins representations rather than adding numeric values. Formatting should occur after the correct numerical result is computed.
Question 55
An agent can prepare a deployment plan, but a named release owner must approve execution. Which state transition should the application enforce?
- Keep preparation separate from execution and require recorded owner approval.
- Treat a well-written plan as the owner’s approval.
- Let the planning agent impersonate the release owner.
- Deploy first and request approval if monitoring detects trouble.
Explanation
Correct Answer
A:Planning authority is not deployment authority. An explicit approval transition prevents the agent’s completed plan from being mistaken for permission to execute it.
Incorrect Answers
B:Document quality does not identify the authorized approver or their decision. The application would be inventing permission from presentation.
C:Role wording does not confer another person’s authority. Approval must come through the actual authorized identity and workflow.
D:That reverses the required order. Monitoring can detect problems, but it cannot retroactively satisfy a prerequisite approval.
Question 56
A repair agent’s plan assumes a service is stopped, but another operator may restart it before execution. What should happen before the dependent repair step?
- Increase confidence in the original observation.
- Hide the state check from the execution trace.
- Recheck the relevant live state and reconcile any changed precondition.
- Assume the plan freezes external service state.
Explanation
Correct Answer
C:A plan is based on observations that can become stale. Revalidating the precondition prevents execution from relying on a state that no longer exists.
Incorrect Answers
A:Confidence does not establish freshness. A previously accurate observation can become obsolete without having been uncertain when collected.
B:Removing evidence makes diagnosis harder and does not prevent changes. The application needs current state, not less visibility into it.
D:Planning does not lock another system unless a real coordination mechanism does so. Other operators can still change the service.
Question 57
Claude requests an unregistered client tool name. What should the harness do?
- Report success without running anything.
- Map it silently to the closest destructive tool.
- Evaluate the name as a shell command.
- Reject the unknown operation through a controlled error path.
Explanation
Correct Answer
D:Dispatch must use an explicit registry. An unrecognized model-produced name must not become arbitrary executable code.
Incorrect Answers
A:A fabricated success gives subsequent reasoning a false observation and conceals the unresolved operation.
B:Approximate naming does not establish intent; a mistaken mapping can execute a materially different operation.
C:This turns generated text into unrestricted execution and bypasses the intended tool boundary.
Question 58
A client tool requires an integer quantity, but receives an object. What must happen before inventory changes?
- Run the operation and validate its result afterward.
- Convert every object to the number one.
- Assume model-generated JSON satisfies every tool contract.
- Validate the arguments and return a useful validation failure.
Explanation
Correct Answer
D:The execution boundary must reject incompatible inputs before mutation. The model can then correct the request using explicit feedback.
Incorrect Answers
A:Post-execution validation cannot prevent a mutation already performed using invalid arguments.
B:Silent coercion invents a quantity and can cause incorrect inventory changes without revealing the malformed request.
C:JSON syntax does not guarantee agreement with the declared parameter types or business constraints.
Question 59
A user approves deleting file A. Before execution, the agent changes the target to file B. What should the harness require?
- Approval bound to the actual target and operation being executed.
- Reuse approval because both actions involve deletion.
- Treat any earlier approval as session-wide permission.
- Ask the model whether the change seems harmless.
Explanation
Correct Answer
A:The previous approval covers a different action. Binding approval to normalized execution parameters prevents silent expansion of its scope.
Incorrect Answers
B:The operation category alone omits the approved resource and therefore does not authorize the changed target.
C:A narrow approval is not a general grant for every later action in the session.
D:The model’s assessment cannot replace the user’s decision about the specific resource to delete.
Question 60
The application enables an Anthropic-hosted web search tool. Who performs that tool’s actual search execution?
- Anthropic’s server-side tool infrastructure.
- The application’s JSON serializer.
- The browser displaying the assistant’s final text.
- A custom function automatically generated in the application.
Explanation
Correct Answer
A:Server tools execute on the provider’s infrastructure. The application still handles returned content and applicable completion conditions.
Incorrect Answers
B:Serialization prepares request data; it does not implement a search engine or execute hosted tools.
C:Rendering results does not make the browser responsible for the API’s server-side search operation.
D:Declaring a server tool does not create a local handler or install generated function code.