Free CCAR-F Agentic Architecture & Orchestration Practice Questions

The Agentic Architecture & Orchestration domain makes up 27% of the Anthropic CCAR-F exam. Practice free questions covering agent loops, workflows vs agents, orchestrator–worker patterns, subagent delegation, and multi-agent system design — each with the correct answer and a detailed explanation.

CCAR-F Agentic Architecture & Orchestration Sample Questions with Answers

Sample Question 1 — Agentic Architecture & Orchestration

A support agent calls several account tools before composing a customer response. The current application stops whenever Claude's text contains the word "completed," causing it to terminate before some requested tools run. Which implementation BEST improves reliability?

  1. A. Continue until two consecutive responses contain no tool names.
  2. B. Execute requests on tool_use, append matching tool results, and continue until end_turn. (Correct answer)
  3. C. Begin a new conversation after each tool call to reduce context size.
  4. D. Require Claude to end every final response with a unique completion phrase.

Correct answer: B

Explanation: Correct answer (B): The application should treat stop_reason as the protocol-level control signal. On tool_use, it executes the requested tool, appends the correctly associated tool result to conversation history, and sends the updated conversation back to Claude. It terminates the current loop on end_turn. This matters in production because free-form wording is variable, whereas protocol-level signals and preserved tool history support reliable multi-step execution. Why the other options are wrong: - Option A: Tool names in generated text are not authoritative execution signals, and waiting for two responses could add unnecessary latency. - Option C: Fresh conversations discard the tool requests and prior observations needed to continue the task coherently. - Option D: A completion phrase still relies on nondeterministic natural-language compliance rather than the API's stop reason.

Sample Question 2 — Agentic Architecture & Orchestration

A research agent can search regulatory filings, query market data, and retrieve internal reports. The correct next source depends on evidence returned by the previous tool, but production policy requires a maximum of eight tool calls. Which orchestration approach is MOST appropriate?

  1. A. Always call the three tools in a fixed order and stop after one pass.
  2. B. Let Claude invoke tools without retaining results or enforcing a call limit.
  3. C. Let Claude select each next tool while the host preserves results and enforces the limit. (Correct answer)
  4. D. Ask Claude to list every tool it might need before the host executes any.

Correct answer: C

Explanation: Correct answer (C): Model-driven tool selection is appropriate because each observation can change the next research action. The host should retain every tool request and result, continue according to stop_reason, and enforce the deterministic eight-call limit. This combines adaptive reasoning with application-controlled safety, preventing unbounded production loops while avoiding a rigid sequence. Why the other options are wrong: - Option A: A fixed sequence can suit stable workflows, but it cannot adapt when returned evidence indicates a different source or follow-up. - Option B: Discarding results breaks continuity, and the missing host-enforced limit violates the stated production policy. - Option D: Up-front planning can be useful, but the complete tool sequence cannot be known reliably before the agent observes intermediate evidence.

Sample Question 3 — Agentic Architecture & Orchestration

An investment research system has specialists for patents, financial statements, and industry news. Investigations vary by company, findings can conflict, and the final report must synthesize evidence across sources. Which architecture BEST meets these requirements?

  1. A. Use a coordinator to select specialists, aggregate results, and request targeted refinement. (Correct answer)
  2. B. Have every specialist exchange unrestricted messages with every other specialist until consensus.
  3. C. Run every specialist once and concatenate their responses in a predefined section order.
  4. D. Assign the entire investigation to one specialist selected from the company industry code.

Correct answer: A

Explanation: 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.

Sample Question 4 — Agentic Architecture & Orchestration

A repository-wide authentication change affects an API handler, a shared token library, integration tests, and deployment documentation. The handler depends on the revised library interface, while the documentation and initial test analysis can proceed independently. Which decomposition strategy is BEST?

  1. A. Give each subagent several isolated code lines, then merge all edits by timestamp.
  2. B. Let all subagents modify every affected file concurrently and resolve conflicts automatically.
  3. C. Complete all code, tests, and documentation serially with one general-purpose subagent.
  4. D. Assign coherent concerns, sequence interface-dependent work, parallelize independent work, then run integration review. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): The coordinator should divide work into coherent concerns, provide the necessary interface context, sequence the handler work after the shared-library contract is known, and parallelize genuinely independent analysis. A final integration review can detect inconsistencies across code, tests, and documentation. In production, dependency-aware orchestration reduces conflicting edits and catches cross-artifact omissions. Why the other options are wrong: - Option A: Arbitrary line-level fragments deprive agents of coherent context and make integration unreliable. - Option B: Overlapping concurrent edits create avoidable conflicts and ignore the stated interface dependency. - Option C: Serial execution avoids merge conflicts but sacrifices useful parallelism and specialist boundaries.

Sample Question 5 — Agentic Architecture & Orchestration

A Claude-based coding coordinator has an AgentDefinition for a dependency-audit subagent, but the coordinator cannot launch it. The coordinator's allowedTools contains Read and Grep but not Task. The subagent also needs the target package path and approved-license list. What should the architect change?

  1. A. Add Bash to the coordinator and let the subagent discover all requirements.
  2. B. Add Task to allowedTools and pass the package path and license constraints explicitly. (Correct answer)
  3. C. Move the subagent instructions into the coordinator's system prompt and remove its definition.
  4. D. Grant every available tool to both agents and rely on repository documentation for context.

Correct answer: B

Explanation: Correct answer (B): The coordinator needs the Task tool to invoke the defined subagent, and relevant task context should be passed explicitly. The subagent should receive only the tools needed for dependency auditing. This matters in production because explicit context prevents incorrect assumptions, while least-privilege configuration reduces operational risk. Why the other options are wrong: - Option A: Adding Bash does not provide the Task capability needed to launch the subagent, and discovery may miss required constraints. - Option C: This discards the intended specialist boundary and does not fix subagent invocation. - Option D: Broad tool access violates least privilege and still assumes that required context will be inferred.

Sample Question 6 — Agentic Architecture & Orchestration

A support agent may issue refunds above $1,500 only after identity verification and policy approval. Skipping either step creates material financial risk. If approval is unavailable, a human reviewer must receive a complete handoff. Which design BEST satisfies these requirements?

  1. A. Tell Claude to verify and approve first, then trust its generated explanation as evidence.
  2. B. Allow the refund immediately, but send the transaction to a reviewer for later confirmation.
  3. C. Enforce both checks before execution and hand off completed checks, facts, reason, and unresolved issues. (Correct answer)
  4. D. Ask the customer to confirm twice, then treat the repeated confirmation as policy approval.

Correct answer: C

Explanation: Correct answer (C): Identity verification and policy approval are mandatory high-risk prerequisites, so application logic must block refund execution until both succeed. If approval is unavailable, the workflow should escalate with a structured record of the reason, completed checks, relevant facts, and unresolved issues. Deterministic gates prevent prompt noncompliance from causing irreversible production actions. Why the other options are wrong: - Option A: Prompt instructions and generated explanations do not guarantee that required checks occurred or provide authoritative approval. - Option B: Post-action review occurs too late when an unauthorized high-value refund has already executed. - Option D: Repeated customer confirmation does not replace an independent policy approval requirement.

Sample Question 7 — Agentic Architecture & Orchestration

A support agent uses a payment tool whose amount field may be returned as "$42.10," "4210 cents," or a numeric decimal. Refund calls must also be blocked when the verified account ID differs from the requested account ID. Which implementation is MOST appropriate?

  1. A. Add examples to the prompt and ask Claude to check account IDs and normalize amounts.
  2. B. Normalize only in the reporting database and audit account mismatches at the end of each day.
  3. C. Let each downstream subagent interpret the amount and independently decide whether to refund.
  4. D. Intercept mismatched calls before execution and normalize successful results in PostToolUse. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): The account mismatch must be intercepted at the controlled pre-execution boundary so the prohibited refund never runs. A PostToolUse hook can then convert successful tool responses into a stable amount schema before subsequent reasoning. Hooks provide deterministic enforcement and normalization, which reduces inconsistent decisions and prevents policy violations in production. Why the other options are wrong: - Option A: Prompt examples do not provide a deterministic control for blocking financially sensitive calls. - Option B: Delayed auditing neither protects the live refund boundary nor gives the agent normalized data for subsequent reasoning. - Option C: Duplicated parsing and policy decisions create inconsistent behavior across downstream agents.

Sample Question 8 — Agentic Architecture & Orchestration

A production defect initially appears in invoice_writer.py, but early inspection may reveal dependencies in schema adapters, retry logic, or database migrations. The full investigation scope is unknown, and the final diagnosis must explain cross-file interactions. Which strategy is BEST?

  1. A. Use a fixed chain that always analyzes the same four files in the same order.
  2. B. Analyze relevant files locally, adapt the plan as dependencies emerge, then perform cross-file synthesis. (Correct answer)
  3. C. Send the entire repository to several agents and accept the first diagnosis returned.
  4. D. Analyze invoice_writer.py only because expanding scope could increase latency and token use.

Correct answer: B

Explanation: Correct answer (B): Dynamic decomposition fits an investigation whose required steps depend on newly discovered evidence. Focused per-file analysis controls local complexity and can exploit safe parallelism, while a final cross-file pass explains interfaces and interactions. In production, adaptive planning avoids both rigid blind spots and unstructured repository-wide analysis. Why the other options are wrong: - Option A: A fixed chain is appropriate when scope and dependencies are stable, but the relevant files here are unknown at the outset. - Option C: Duplicating the entire repository review lacks deliberate decomposition, and accepting the first response provides no integration assurance. - Option D: Refusing to expand the investigation after dependencies emerge prevents a reliable system-level diagnosis.

Keep Practicing

Take the 10-question CCAR-F quick-start test across all 5 domains, or return to the CCAR-F practice test hub for 250+ questions, exam details, and a study plan. Drill the other domains: Agentic Architecture & Orchestration, Claude Code Configuration & Workflows, Prompt Engineering & Structured Output, Tool Design & MCP Integration, and Context Management & Reliability.