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.
Question 61
A team wants the provider to operate the agent loop and session infrastructure. Which offering should it evaluate first?
- A local interactive terminal with no service process.
- Only a bare HTTP client calling Messages.
- Claude Managed Agents.
- Only a repository instruction file.
Explanation
Correct Answer
C:This surface provides managed agent orchestration. The team should then verify its required environment, tools, and deployment controls.
Incorrect Answers
A:An interactive terminal does not satisfy the requirement for provider-operated agent infrastructure.
B:Direct model calls leave orchestration and conversation-state management to the application rather than operating them as a managed agent service.
D:Instructions guide behavior; they do not supply runtime hosting, session infrastructure, or execution services.
Question 62
An agent may run locally while calling Claude remotely. Which statement correctly separates these responsibilities?
- Remote inference means every custom tool runs remotely too.
- Local execution implies the model weights run locally.
- Tool execution location and model inference location are distinct choices.
- The tool’s name determines its physical hosting region.
Explanation
Correct Answer
C:A local harness can execute tools and send model requests remotely. Deployment diagrams must show both boundaries explicitly.
Incorrect Answers
A:Client tools can remain in the application environment even when inference uses a remote API.
B:Running the harness locally does not establish where model inference occurs or place model weights on the machine.
D:Names describe capabilities; hosting and data-routing configuration determine where execution and processing occur.
Question 63
A user cancels a long agent task while a tool call is active. What should the application’s cancellation design address?
- Stop new work and handle the active operation’s cancellation or reconciliation.
- Assume cancellation reverses every completed operation.
- Delete the trace before stopping the worker.
- Hide the interface while continuing all actions.
Explanation
Correct Answer
A:Cancellation must cover orchestration and in-flight work. Some effects may already have occurred and require checking rather than assuming rollback.
Incorrect Answers
B:A cancellation signal is not a transaction rollback across external services.
C:Removing evidence does not terminate execution and makes outstanding effects harder to reconcile.
D:Hiding progress does not cancel execution and can leave the user unaware of continuing side effects.
Question 64
A repair agent says “finished,” but no required test result exists. What should determine application-level completion?
- The explicit acceptance checks and recorded evidence.
- The absence of further tool calls.
- The number of tokens generated.
- The presence of the word “finished.”
Explanation
Correct Answer
A:A completion statement is an observation from the model, not proof that mandatory checks passed. The harness should verify the task contract.
Incorrect Answers
B:An agent can stop prematurely. Inactivity does not demonstrate that every acceptance condition has been satisfied.
C:Output volume does not establish whether the repair works or the required tests executed.
D:A text phrase can be generated without the required execution evidence and is therefore an insufficient completion signal.
Question 65
A code agent needs one project directory. Which filesystem setup minimizes unnecessary access?
- Mount all user directories for convenience.
- Expose the required project paths with only necessary write permissions.
- Make the whole filesystem writable to avoid permission errors.
- Use a prompt to hide sensitive directories after mounting them.
Explanation
Correct Answer
B:Narrow mounts reduce the reachable filesystem. Read and write scope should follow the task rather than the process owner’s entire home directory.
Incorrect Answers
A:Broad mounting exposes unrelated files and secrets even though the task needs only one project.
C:This removes a useful boundary and grants mutation rights beyond the task’s requirements.
D:A textual reminder does not remove filesystem authority or prevent an execution tool from reaching mounted paths.
Question 66
An agent image passes staging tests. Production pulls the tag latest, which now points to another build. The dependency lockfile has not changed. What prevents this artifact substitution?
- Record the image's display name without its digest.
- Deploy the tested image by its immutable digest.
- Pull latest immediately before every agent run.
- Rebuild production from the lockfile and assume the output is identical.
Explanation
Correct Answer
B:A lockfile describes dependency resolution, but an image also includes application files and build outputs. A digest identifies the exact artifact validated in staging instead of a mutable tag target.
Incorrect Answers
A:A human-readable name can refer to multiple builds over time. The record lacks the immutable identity needed to detect this substitution.
C:Frequent pulling increases exposure to tag changes. It does not bind a production run to the artifact that passed validation.
D:The rebuild can differ in source, base image, build tooling, or generated files. Matching dependency versions alone does not prove byte-identical deployment artifacts.
Question 67
A tool can wait indefinitely for an upstream response. Which control belongs in its execution wrapper?
- An explicit timeout and a defined timeout result.
- An unlimited output-token setting.
- A success response issued when the timeout starts.
- A longer tool description.
Explanation
Correct Answer
A:The wrapper must bound waiting and communicate failure accurately. Otherwise one stalled dependency can prevent the agent from progressing or terminating.
Incorrect Answers
B:Generation limits do not terminate a network operation that is blocked inside the tool implementation.
C:Starting a timer is not successful execution. Reporting success before an outcome corrupts the agent’s evidence.
D:Descriptions influence selection but do not enforce timeouts on the underlying network connection.
Question 68
A hosted agent runs two repositories in separate jobs. A tool writes to a relative path named output.json. Which runtime setting prevents the file from landing in the wrong project?
- Give each job an explicit working directory within its assigned workspace.
- Use the same output filename and assume job IDs change filesystem resolution.
- Let all jobs inherit whichever directory the last job used.
- Include the repository name only in the final response.
Explanation
Correct Answer
A:Relative paths resolve against the process working directory; explicit per-job configuration keeps file operations associated with the intended repository.
Incorrect Answers
B:A job identifier does not affect relative paths unless the runtime explicitly maps it to an isolated directory.
C:Shared mutable process state can redirect later operations into another job's workspace.
D:A descriptive label does not change how the operating system resolves the tool's relative path.
Question 69
One assistant turn requests two independent client tools. Which TWO practices preserve correct result handling? Choose TWO.
- Attach both results to whichever call finished first.
- Discard a slower result because another tool succeeded.
- Return all required results in the supported next-message structure.
- Track each result with its originating tool-use identifier.
- Replace both tool requests with a generic success sentence.
Explanation
Correct Answers
C:A complete structured handoff allows the model to reason from both observations without leaving an invocation unresolved.
D:Identifiers associate observations with the exact requests, even when operations complete in different orders.
Incorrect Answers
A:Completion order does not change request identity; combining identifiers loses the second call’s correspondence.
B:Independent calls may supply different required facts. One success does not make the other observation unnecessary.
E:Unstructured summary text loses the protocol associations needed to represent the completed invocations accurately.
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.