CCAR-F Practice Questions: Agentic Architecture & Orchestration Domain

Correct answer (A): The application should use protocol-level stop reasons: tool_use means it must execute the requested tool and append the result before calling Claude again, while end_turn normally signals completion. This is more deterministic than parsing generated language. In production, protocol-driven loop control prevents premature termination and preserves the observations Claude needs for subsequent reasoning.

Why the other options are wrong:
- Option B: The absence of tool names in generated text does not reliably indicate whether Claude requested a tool through the API protocol.
- Option C: A unique phrase is easier to parse than a common word, but it still depends on probabilistic natural-language compliance.
- Option D: Calling every tool adds latency and cost and may invoke irrelevant tools instead of following Claude's actual requests.

Correct answer (C): A hub-and-spoke pattern gives a coordinator responsibility for dynamic decomposition, specialist selection, and final synthesis. Specialists receive focused assignments, while the coordinator reconciles their findings into a coherent answer. This matters in production because it limits unnecessary work, provides a clear control point, and handles cross-cutting conclusions better than simple concatenation.

Why the other options are wrong:
- Option A: Invoking all specialists wastes resources when expertise needs vary, and returning one response unchanged does not reconcile evidence.
- Option B: Peer delegation may support emergent collaboration, but it weakens control over scope, completion, and responsibility for synthesis.
- Option D: A fixed decomposition is predictable, but it does not accommodate requests whose evidence and expertise needs differ.

Correct answer (B): The model needs the conversation history that established the objective, the assistant message containing the tool request, and the tool result associated with that request. Supplying only a raw tool output breaks the interaction sequence and removes task intent. Preserving complete tool exchanges matters in production because autonomous tasks commonly require several related tool cycles.

Why the other options are wrong:
- Option A: Reducing result volume may help manage context, but it does not restore the missing objective or tool-call relationship.
- Option C: Tools can return structured records for Claude to interpret; they do not need to produce the final user-facing answer.
- Option D: A one-tool limit would prematurely terminate research tasks that legitimately require multiple tool calls.

Correct answer (B): Dividing work by coherent concerns gives each specialist enough context to reason meaningfully, while a coordinator-led integration pass examines interactions across components and requests refinement where needed. The underlying principle is that decomposition boundaries must preserve cross-cutting behavior rather than maximize task count. In production, this retains parallelism without systematically overlooking defects that emerge only from component interactions.

Why the other options are wrong:
- Option A: Finer decomposition may increase concurrency, but it further removes the context needed to discover cross-component authorization behavior.
- Option C: A repository-wide agent may see more interactions, but abandoning all decomposition can overwhelm context and sacrifices useful specialization and parallelism.
- Option D: Confidence filtering can prioritize findings, but it cannot detect a flaw that no isolated method-level agent had enough context to identify.

Correct answer (A): Keeping tool requests and their corresponding results in the same conversation gives Claude direct access to accumulated findings when selecting the next action. The underlying principle is that adaptive agentic execution depends on preserving relevant state across iterations, although verbose results may be normalized without losing material evidence. In production, continuity reduces duplicated work and makes follow-up research responsive to actual discoveries.

Why the other options are wrong:
- Option B: Batching fixed queries can provide broad coverage, but it removes the ability to adapt later searches to earlier evidence.
- Option C: Longer summaries might retain more information, but they do not reliably preserve the structured relationship between tool requests and results.
- Option D: Multiple services may diversify sources, but they do not address the orchestrator's loss of conversational state.

Correct answer (A): The application should use the API's stop reasons as the control signal. When Claude returns tool_use, it should execute the requested tool and preserve both the assistant tool request and its associated result before invoking Claude again. An end_turn normally identifies the completed response. This protocol-based loop is more reliable in production than interpreting natural-language completion phrases.

Why the other options are wrong:
- Option B: Tool-name and completion-phrase parsing is brittle because natural-language wording can vary.
- Option C: Executing every tool in advance defeats model-directed tool selection and may perform unnecessary or unsafe actions.
- Option D: Discarding the tool-request message and starting a new conversation loses context and can cause repetition or inconsistent reasoning.

Correct answer (C): Because the prerequisites are mandatory and the action is consequential, application logic or the tool authorization boundary must make the refund unavailable until both checks pass. Prompts can guide sequencing, but they are not deterministic controls. Production enforcement prevents an occasional model deviation from becoming a financial or compliance incident.

Why the other options are wrong:
- Option A: Repeated instructions may improve adherence but cannot guarantee enforcement under every input and execution path.
- Option B: A post-execution review occurs too late to prevent an unauthorized refund.
- Option D: Lower temperature and explicit planning may reduce variation but do not enforce prerequisites at the authorization boundary.

Correct answer (B): Hooks provide an application-controlled boundary for deterministic tool-policy interception, while post-execution processing can normalize inconsistent tool results before further model reasoning. The underlying principle is that mandatory controls and stable interfaces belong in deterministic orchestration layers rather than relying solely on model interpretation. In production, this prevents unauthorized actions and reduces reasoning errors caused by inconsistent downstream data.

Why the other options are wrong:
- Option A: Claude may normalize values and follow limits, but prompt-only handling does not deterministically enforce the settlement policy.
- Option C: Examples can improve format interpretation, but retrying smaller settlements could evade policy intent and still lacks deterministic normalization.
- Option D: Separate normalization calls do not intercept settlements above the authorized limit and unnecessarily fragment the workflow.

Correct answer (B): Per-file tasks preserve local detail and manageable scope, while a cross-file integration pass detects mismatched interfaces and dependencies. Adaptive follow-ups allow the workflow to investigate issues discovered during review rather than applying unnecessary checks everywhere. This hybrid decomposition improves production review coverage without losing system-level consistency.

Why the other options are wrong:
- Option A: Isolated reviews can find local defects but cannot reliably detect incompatibilities that emerge across file or package boundaries.
- Option C: A single agent retains broad context, but prohibiting decomposition makes a large review harder to focus and limits targeted parallel analysis.
- Option D: Service and adapter implementations can violate shared contracts even when the type definitions themselves appear correct.

Correct answer (A): A hub-and-spoke coordinator can decompose the investigation, choose only relevant specialists, aggregate their focused findings, and request refinement when evidence conflicts or remains incomplete. This architectural pattern matters in production because it centralizes synthesis and limits coordination complexity while preserving adaptive specialist use.

Why the other options are wrong:
- Option B: Unrestricted all-to-all communication increases coordination complexity and weakens ownership of final synthesis.
- Option C: One-pass execution and concatenation do not reconcile contradictions or support iterative refinement.
- Option D: A single specialist cannot reliably integrate evidence spanning patents, finances, and industry developments.