Question 120
A team maintaining an internal agent-extension rollout has two different engineering requirements: (1) expose an MCP server to clients across a network boundary, and (2) offer users a reusable parameterized instruction template from an MCP server. Which TWO independent changes directly meet them? Choose TWO.
- Register a prompt template with its required arguments.
- Expose the manual as an MCP resource rather than a tool.
- Configure the supported tool_choice behavior for the request.
- Validate model-supplied tool arguments against schema and domain rules.
- Configure a supported remote HTTP transport with authentication and deployment controls.
Explanation
Correct Answers
A:To offer users a reusable parameterized instruction template from an MCP server as part of a design that must also expose an MCP server to clients across a network boundary, MCP prompts provide structured prompt content, but by themselves do not perform external business actions; this is why the action is one of the two required selections.
E:To expose an MCP server to clients across a network boundary as part of a design that must also offer users a reusable parameterized instruction template from an MCP server, remote connectivity needs network routing, identity, and authorization; a local stdio subprocess is not a remotely reachable service; this is why the action is one of the two required selections.
Incorrect Answers
B:This alternative serves the separate goal to supply a read-only policy manual to an MCP client without giving it an action endpoint because a resource is contextual data; representing a passive document as an imperative function adds unnecessary execution semantics; the required outcomes are instead to expose an MCP server to clients across a network boundary and to offer users a reusable parameterized instruction template from an MCP server.
C:This alternative serves the separate goal to require or restrict tool use for a specific turn when the API supports that control because tool-selection controls influence whether Claude may call tools, but cannot grant permissions to the underlying systems; the required outcomes are instead to expose an MCP server to clients across a network boundary and to offer users a reusable parameterized instruction template from an MCP server.
D:This alternative serves the separate goal to reject a date or identifier that violates business constraints before a service call because even syntactically correct generated inputs can violate account scope or allowed values, so execution must validate both levels; the required outcomes are instead to expose an MCP server to clients across a network boundary and to offer users a reusable parameterized instruction template from an MCP server.
Question 121
The team responsible for a business workflow that calls Claude needs to catch mismatches between model calls, tool handlers, and application persistence. What is the most direct engineering change?
- Release to a small monitored cohort before broad deployment.
- Route Claude requests through a trusted backend service.
- Exercise complete request-to-tool-to-response paths in a controlled environment.
- Inspect response.content blocks and metadata through the SDK's typed representation.
Explanation
Correct Answer
C:To catch mismatches between model calls, tool handlers, and application persistence in this business workflow that calls Claude, unit tests cannot reveal every serialization or permission issue across service boundaries; integration tests validate the connected contract; the proposed implementation directly supplies that capability.
Incorrect Answers
A:The proposed alternative helps to limit the blast radius of a new model configuration during rollout because canary exposure provides real traffic signals with an option to stop or roll back, unlike a simultaneous change for every customer; that mechanism does not meet the different requirement to catch mismatches between model calls, tool handlers, and application persistence.
B:The proposed alternative helps to keep a browser application from exposing the secret used for model calls because a backend controls credentials and access policy while the UI sends user input through an authenticated application endpoint; that mechanism does not meet the different requirement to catch mismatches between model calls, tool handlers, and application persistence.
D:The proposed alternative helps to avoid assuming a structured SDK response is a plain string because the SDK exposes block types, stop reason, and usage separately; naïve string concatenation can lose tool requests or other content; that mechanism does not meet the different requirement to catch mismatches between model calls, tool handlers, and application persistence.
Question 122
For a safety-sensitive Claude application, a design is acceptable only if it can protect customer identity fields when only document topics are needed for classification while also being able to resist instructions embedded in a retrieved support ticket that demand secret disclosure. Which paired implementation is sufficient?
- Perform both steps: (1) Use layered policy enforcement and restrict powerful tools independently of chat text; (2) treat retrieved content as untrusted and prevent its instructions from controlling tools.
- Perform both steps: (1) Use layered policy enforcement and restrict powerful tools independently of chat text; (2) redact secrets and protect log access with retention controls.
- Perform both steps: (1) Remove or mask unnecessary personal information before model submission; (2) treat retrieved content as untrusted and prevent its instructions from controlling tools.
- Perform both steps: (1) Remove or mask unnecessary personal information before model submission; (2) redact secrets and protect log access with retention controls.
Explanation
Correct Answer
C:The first change enables the team to protect customer identity fields when only document topics are needed for classification because minimization limits exposure while preserving task inputs; relying on the model to ignore visible PII does not provide the same protection; the second addresses the separate need to resist instructions embedded in a retrieved support ticket that demand secret disclosure because source material is data rather than authority; enforcing tool permissions and separating roles limits an attacker's ability to redirect execution.
Incorrect Answers
A:The second step meets the requirement to resist instructions embedded in a retrieved support ticket that demand secret disclosure because source material is data rather than authority; enforcing tool permissions and separating roles limits an attacker's ability to redirect execution; the first step would only help to prevent an adversarial user from bypassing safety policy by roleplay instructions because adversarial prompting can target instruction hierarchy, so product controls and tool authorization must remain effective when wording is manipulated, leaving the separate requirement to protect customer identity fields when only document topics are needed for classification unmet.
B:Both steps target different results—prevent an adversarial user from bypassing safety policy by roleplay instructions and investigate user actions without leaking API keys into operational records—because adversarial prompting can target instruction hierarchy, so product controls and tool authorization must remain effective when wording is manipulated and logs support accountability only when access and retention are restricted; copying all prompt content into logs creates another exposure surface; the required capabilities are instead to protect customer identity fields when only document topics are needed for classification and to resist instructions embedded in a retrieved support ticket that demand secret disclosure.
D:The first step is appropriate for the requirement to protect customer identity fields when only document topics are needed for classification because minimization limits exposure while preserving task inputs; relying on the model to ignore visible PII does not provide the same protection; the other step instead aims to investigate user actions without leaking API keys into operational records because logs support accountability only when access and retention are restricted; copying all prompt content into logs creates another exposure surface, which does not meet the second requirement to resist instructions embedded in a retrieved support ticket that demand secret disclosure.
Question 123
An architecture review of a tool-using incident triage agent asks how to stop two parallel branches from overwriting the same mutable record. Which action is the best fit?
- Use a supervisor agent to delegate bounded tasks and synthesize results.
- Compact prior progress into an accurate working summary.
- Define and validate a typed schema at the handoff.
- Use isolated branch state and a controlled reconciliation step.
Explanation
Correct Answer
D:To stop two parallel branches from overwriting the same mutable record in this tool-using incident triage agent, explicit merge rules are required when outputs converge; concurrent mutation without arbitration causes nondeterministic results; the proposed implementation directly supplies that capability.
Incorrect Answers
A:The proposed alternative helps to coordinate several specialized investigations and assemble one accountable final answer because a supervisor owns the overall objective and integrates subagent findings; each delegate should have a narrow mandate and return a constrained summary; that mechanism does not meet the different requirement to stop two parallel branches from overwriting the same mutable record.
B:The proposed alternative helps to continue a long agent run after intermediate observations dominate the context because compaction preserves decisions and unresolved work while reducing token pressure; losing constraints in the summary can still cause errors; that mechanism does not meet the different requirement to stop two parallel branches from overwriting the same mutable record.
C:The proposed alternative helps to pass reliable machine-readable outputs between independently developed agent stages because typed outputs make downstream contracts inspectable and reduce ambiguity caused by free-form prose passed as control data; that mechanism does not meet the different requirement to stop two parallel branches from overwriting the same mutable record.
Question 124
An architecture decision for a TypeScript Claude integration must support send a model inference request with the operation expected by the REST endpoint and authenticate a server-to-server Claude API request. Which option gives a complete rather than partial solution?
- Perform both steps: (1) Set the appropriate application/json content type and serialize a valid payload; (2) use asynchronous sdk calls with bounded concurrency.
- Perform both steps: (1) Set the appropriate application/json content type and serialize a valid payload; (2) place a valid secret in the documented api authentication header.
- Perform both steps: (1) Use the API's documented POST request with a JSON body; (2) use asynchronous sdk calls with bounded concurrency.
- Perform both steps: (1) Use the API's documented POST request with a JSON body; (2) place a valid secret in the documented api authentication header.
Explanation
Correct Answer
D:The first change enables the team to send a model inference request with the operation expected by the REST endpoint because a state-changing inference operation is submitted as an HTTP request with an explicit method and payload, not inferred from a browser URL alone; the second addresses the separate need to authenticate a server-to-server Claude API request because API authentication is a request transport concern; placing a key inside the prompt body neither authenticates the request nor protects it.
Incorrect Answers
A:Both steps target different results—have the server parse a JSON request body predictably and serve many concurrent network-bound Claude requests without blocking worker threads unnecessarily—because a JSON body requires both correct serialization and matching HTTP metadata; raw language-model text cannot replace a valid request object and async I/O can overlap network waits, but it does not remove provider limits or make CPU-heavy postprocessing nonblocking; the required capabilities are instead to send a model inference request with the operation expected by the REST endpoint and to authenticate a server-to-server Claude API request.
B:The second step meets the requirement to authenticate a server-to-server Claude API request because API authentication is a request transport concern; placing a key inside the prompt body neither authenticates the request nor protects it; the first step would only help to have the server parse a JSON request body predictably because a JSON body requires both correct serialization and matching HTTP metadata; raw language-model text cannot replace a valid request object, leaving the separate requirement to send a model inference request with the operation expected by the REST endpoint unmet.
C:The first step is appropriate for the requirement to send a model inference request with the operation expected by the REST endpoint because a state-changing inference operation is submitted as an HTTP request with an explicit method and payload, not inferred from a browser URL alone; the other step instead aims to serve many concurrent network-bound Claude requests without blocking worker threads unnecessarily because async I/O can overlap network waits, but it does not remove provider limits or make CPU-heavy postprocessing nonblocking, which does not meet the second requirement to authenticate a server-to-server Claude API request.
Question 125
A team maintaining a developer workstation running Claude Code has two different engineering requirements: (1) give Claude Code repository-specific build and test instructions for future sessions, and (2) delegate an independent code review without filling the main coding session with details. Which TWO independent changes directly meet them? Choose TWO.
- Add concise project guidance to the appropriate CLAUDE.md file.
- Place a path-targeted rule in the repository's rules configuration.
- Configure a specialized Claude Code subagent with a narrow remit.
- Use the supported session continuation or resume mechanism.
- Invoke the supported noninteractive or print-mode command interface.
Explanation
Correct Answers
A:To give Claude Code repository-specific build and test instructions for future sessions as part of a design that must also delegate an independent code review without filling the main coding session with details, Claude Code reads persistent project instructions from its memory hierarchy; one-off chat messages do not reliably establish durable repository guidance; this is why the action is one of the two required selections.
C:To delegate an independent code review without filling the main coding session with details as part of a design that must also give Claude Code repository-specific build and test instructions for future sessions, subagents can focus on a task and return a summary, limiting noise in the parent session and allowing separate permissions when configured; this is why the action is one of the two required selections.
Incorrect Answers
B:This alternative serves the separate goal to apply a coding convention only when Claude Code edits a particular path because scoped rules keep specialized instructions close to their applicable files, avoiding global instructions that conflict with other modules; the required outcomes are instead to give Claude Code repository-specific build and test instructions for future sessions and to delegate an independent code review without filling the main coding session with details.
D:This alternative serves the separate goal to continue a previously interrupted Claude Code work session because restoring session context reduces lost work while preserving continuity of prior tool activity; opening a new unrelated session does not do the same; the required outcomes are instead to give Claude Code repository-specific build and test instructions for future sessions and to delegate an independent code review without filling the main coding session with details.
E:This alternative serves the separate goal to run a Claude Code task from an automated CI job without an interactive terminal because headless operation outputs results for automation and requires explicit credential, permission, and failure handling in the CI environment; the required outcomes are instead to give Claude Code repository-specific build and test instructions for future sessions and to delegate an independent code review without filling the main coding session with details.
Question 126
A change request for a context-rich document assistant combines two needs: teach a classifier how to label borderline requests consistently, and measure whether a changed instruction actually improves extraction accuracy. Which implementation pair addresses both, not just one?
- Perform both steps: (1) Retain essential fields and discard irrelevant tool output from active context; (2) specify the forbidden content alongside an enforceable validation check.
- Perform both steps: (1) Supply representative labeled examples including edge conditions; (2) specify the forbidden content alongside an enforceable validation check.
- Perform both steps: (1) Retain essential fields and discard irrelevant tool output from active context; (2) hold evaluation inputs constant and compare prompt versions.
- Perform both steps: (1) Supply representative labeled examples including edge conditions; (2) hold evaluation inputs constant and compare prompt versions.
Explanation
Correct Answer
D:The first change enables the team to teach a classifier how to label borderline requests consistently because diverse demonstrations make the intended decision boundary clearer than an isolated easy example, provided labels are accurate; the second addresses the separate need to measure whether a changed instruction actually improves extraction accuracy because a reproducible prompt test isolates the change being studied; anecdotal output from unrelated cases does not show improvement.
Incorrect Answers
A:Both steps target different results—stop verbose tool payloads from consuming most of a long-running agent's context and avoid a prohibited output element such as confidential identifiers—because summarizing or filtering observations keeps the reasoning window focused without deleting required facts from durable application state and negative prompt guidance signals the intended boundary, but a deterministic validator is needed where violating it has consequences; the required capabilities are instead to teach a classifier how to label borderline requests consistently and to measure whether a changed instruction actually improves extraction accuracy.
B:The first step is appropriate for the requirement to teach a classifier how to label borderline requests consistently because diverse demonstrations make the intended decision boundary clearer than an isolated easy example, provided labels are accurate; the other step instead aims to avoid a prohibited output element such as confidential identifiers because negative prompt guidance signals the intended boundary, but a deterministic validator is needed where violating it has consequences, which does not meet the second requirement to measure whether a changed instruction actually improves extraction accuracy.
C:The second step meets the requirement to measure whether a changed instruction actually improves extraction accuracy because a reproducible prompt test isolates the change being studied; anecdotal output from unrelated cases does not show improvement; the first step would only help to stop verbose tool payloads from consuming most of a long-running agent's context because summarizing or filtering observations keeps the reasoning window focused without deleting required facts from durable application state, leaving the separate requirement to teach a classifier how to label borderline requests consistently unmet.
Question 127
A release check on an autonomous repository maintenance job found two unsatisfied conditions: the service must prevent specialists from producing incompatible artifacts for the orchestrator, and it must pinpoint which known workflow stage failed when the final artifact is incomplete. Which complete engineering plan satisfies both?
- Perform both steps: (1) Define structured inputs, expected outputs, and failure states for each delegate; (2) enforce explicit completion checks and a maximum iteration budget.
- Perform both steps: (1) Define structured inputs, expected outputs, and failure states for each delegate; (2) record stage-level traces and intermediate validation outcomes.
- Perform both steps: (1) Consume the agent runtime's streamed event sequence; (2) record stage-level traces and intermediate validation outcomes.
- Perform both steps: (1) Consume the agent runtime's streamed event sequence; (2) enforce explicit completion checks and a maximum iteration budget.
Explanation
Correct Answer
B:The first change enables the team to prevent specialists from producing incompatible artifacts for the orchestrator because explicit delegation contracts make multi-agent integration testable; free-form handoffs often obscure whether a child finished its assigned task; the second addresses the separate need to pinpoint which known workflow stage failed when the final artifact is incomplete because deterministic stages are operationally useful when their boundaries are observable; a final answer alone cannot reveal which step produced the fault.
Incorrect Answers
A:The first step is appropriate for the requirement to prevent specialists from producing incompatible artifacts for the orchestrator because explicit delegation contracts make multi-agent integration testable; free-form handoffs often obscure whether a child finished its assigned task; the other step instead aims to stop iterative tool use when a success condition or resource budget is reached because bounded loops avoid unending tool calls and uncontrolled spend while making incomplete work visible instead of silently continuing, which does not meet the second requirement to pinpoint which known workflow stage failed when the final artifact is incomplete.
C:The second step meets the requirement to pinpoint which known workflow stage failed when the final artifact is incomplete because deterministic stages are operationally useful when their boundaries are observable; a final answer alone cannot reveal which step produced the fault; the first step would only help to deliver intermediate agent events to the client as they occur because event-driven consumption lets the application display progress and inspect tool actions without waiting for the entire multi-step run to finish, leaving the separate requirement to prevent specialists from producing incompatible artifacts for the orchestrator unmet.
D:Both steps target different results—deliver intermediate agent events to the client as they occur and stop iterative tool use when a success condition or resource budget is reached—because event-driven consumption lets the application display progress and inspect tool actions without waiting for the entire multi-step run to finish and bounded loops avoid unending tool calls and uncontrolled spend while making incomplete work visible instead of silently continuing; the required capabilities are instead to prevent specialists from producing incompatible artifacts for the orchestrator and to pinpoint which known workflow stage failed when the final artifact is incomplete.
Question 128
Two acceptance tests block a production inference application: one requires the ability to state how much interruption a customer-facing AI feature can tolerate; the other requires the ability to design for a documented burst of simultaneous requests without uncontrolled overload. Which plan clears both tests?
- Perform both steps: (1) Define availability objectives with failure and recovery criteria; (2) enforce per-tenant token and request budgets in the backend.
- Perform both steps: (1) Track latency, failure rates, token use, and output-quality signals over time; (2) enforce per-tenant token and request budgets in the backend.
- Perform both steps: (1) Define availability objectives with failure and recovery criteria; (2) specify expected concurrency and enforce admission or queueing limits.
- Perform both steps: (1) Track latency, failure rates, token use, and output-quality signals over time; (2) specify expected concurrency and enforce admission or queueing limits.
Explanation
Correct Answer
C:The first change enables the team to state how much interruption a customer-facing AI feature can tolerate because reliability requirements determine fallback, retries, health monitoring, and degraded behavior; they are distinct from the response's linguistic quality; the second addresses the separate need to design for a documented burst of simultaneous requests without uncontrolled overload because capacity planning separates average traffic from peak demand and prevents upstream rate limits from becoming cascading application failures.
Incorrect Answers
A:The first step is appropriate for the requirement to state how much interruption a customer-facing AI feature can tolerate because reliability requirements determine fallback, retries, health monitoring, and degraded behavior; they are distinct from the response's linguistic quality; the other step instead aims to keep one customer from consuming an entire shared inference budget because application quotas isolate noisy tenants and control spending; provider-wide rate limits alone do not guarantee fairness between customers, which does not meet the second requirement to design for a documented burst of simultaneous requests without uncontrolled overload.
B:Both steps target different results—detect rising inference errors after a feature reaches production and keep one customer from consuming an entire shared inference budget—because production health spans infrastructure and model behavior, so a healthy HTTP server alone does not prove the AI feature is functioning and application quotas isolate noisy tenants and control spending; provider-wide rate limits alone do not guarantee fairness between customers; the required capabilities are instead to state how much interruption a customer-facing AI feature can tolerate and to design for a documented burst of simultaneous requests without uncontrolled overload.
D:The second step meets the requirement to design for a documented burst of simultaneous requests without uncontrolled overload because capacity planning separates average traffic from peak demand and prevents upstream rate limits from becoming cascading application failures; the first step would only help to detect rising inference errors after a feature reaches production because production health spans infrastructure and model behavior, so a healthy HTTP server alone does not prove the AI feature is functioning, leaving the separate requirement to state how much interruption a customer-facing AI feature can tolerate unmet.
Question 129
A team maintaining an enterprise agent deployment has two different engineering requirements: (1) require a person to approve an irreversible administrative change, and (2) prevent generated advice that violates the product's publication rules from being shown. Which TWO independent changes directly meet them? Choose TWO.
- Use layered policy enforcement and restrict powerful tools independently of chat text.
- Validate or moderate final content before release.
- Remove or mask unnecessary personal information before model submission.
- Isolate external search results and validate any resulting tool action independently.
- Pause execution for explicit approval at the high-impact action boundary.
Explanation
Correct Answers
B:To prevent generated advice that violates the product's publication rules from being shown as part of a design that must also require a person to approve an irreversible administrative change, output checks catch issues emerging during generation; they complement rather than replace input controls and tool restrictions; this is why the action is one of the two required selections.
E:To require a person to approve an irreversible administrative change as part of a design that must also prevent generated advice that violates the product's publication rules from being shown, human approval is a policy-enforced transition, not a note written in the model's plan or a message displayed afterward; this is why the action is one of the two required selections.
Incorrect Answers
A:This alternative serves the separate goal to prevent an adversarial user from bypassing safety policy by roleplay instructions because adversarial prompting can target instruction hierarchy, so product controls and tool authorization must remain effective when wording is manipulated; the required outcomes are instead to require a person to approve an irreversible administrative change and to prevent generated advice that violates the product's publication rules from being shown.
C:This alternative serves the separate goal to protect customer identity fields when only document topics are needed for classification because minimization limits exposure while preserving task inputs; relying on the model to ignore visible PII does not provide the same protection; the required outcomes are instead to require a person to approve an irreversible administrative change and to prevent generated advice that violates the product's publication rules from being shown.
D:This alternative serves the separate goal to reject a malicious instruction inserted into a web-search result rather than the user's direct request because indirect attacks exploit lower-trust retrieval surfaces, so trusting search snippets as operational directives crosses a security boundary; the required outcomes are instead to require a person to approve an irreversible administrative change and to prevent generated advice that violates the product's publication rules from being shown.
Question 130
For a model-output validation layer, the missing capability is to avoid treating fluent model wording as evidence that a claim is verified. What should the developers implement?
- Place essential constraints prominently and structure large evidence clearly.
- Ground consequential statements in retrieved evidence or external checks.
- Summarize prior decisions and preserve unresolved constraints.
- Instruct the application to request clarification for unresolved high-impact ambiguity.
Explanation
Correct Answer
B:To avoid treating fluent model wording as evidence that a claim is verified in this model-output validation layer, linguistic confidence is not a calibrated truth guarantee, so factual correctness requires independent support; the proposed implementation directly supplies that capability.
Incorrect Answers
A:The proposed alternative helps to reduce the chance that critical instructions become buried in a huge prompt because ordering, headings, and explicit references help the model distinguish governing rules from supporting material in large requests; that mechanism does not meet the different requirement to avoid treating fluent model wording as evidence that a claim is verified.
C:The proposed alternative helps to maintain the current task after older dialogue exceeds the active token budget because a faithful rolling summary carries forward objectives and commitments while avoiding a full-history replay on each call; that mechanism does not meet the different requirement to avoid treating fluent model wording as evidence that a claim is verified.
D:The proposed alternative helps to ask for missing essential parameters rather than inventing a consequential choice because clarification reduces irreversible guesses when requirements do not uniquely determine a safe action; that mechanism does not meet the different requirement to avoid treating fluent model wording as evidence that a claim is verified.