Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 2

Question 10

A developer is designing a tool-use loop that may encounter intermittent failures. Which state must be retained to resume safely after a process restart?

  1. Only the user's most recent question
  2. Only the assistant's most optimistic final draft
  3. Only a count of generated text tokens
  4. Tool-use IDs, execution status and conversation state

Explanation

Correct Answer

D:

A durable checkpoint distinguishes attempted side effects from unexecuted calls. That prevents replaying writes and lets follow-up tool results attach to the correct requests.

Incorrect Answers

A:

The current prompt alone cannot reveal which tool operations already ran. Restarts could duplicate billing or produce context-free replies.

B:

A speculative draft does not certify which external operations completed. Persisting it without tool state encourages false completion claims.

C:

Token totals provide no mapping to actions and outcomes. Execution recovery requires the actual call state rather than a usage statistic.

Question 11

An agent sends tool_result data and then immediately displays the prior assistant's pending text as final, without requesting continuation. What step is missing?

  1. Remove tool_result blocks because they are only internal diagnostics
  2. Request another assistant turn using the tool result in conversation history
  3. Increase the temperature so a final answer appears automatically
  4. Rewrite the tool_use id into the user's ticket number

Explanation

Correct Answer

B:

The model needs to consume the returned observation to interpret success or failure and formulate a grounded final response.

Incorrect Answers

A:

Results are essential evidence in the client-tool protocol. Deleting them would leave the model unaware of the operation's actual outcome.

C:

Sampling settings affect generation style, not whether the host has asked the API to continue after receiving a result.

D:

Call identifiers serve protocol correlation, not domain data. Changing them would break matching and cannot substitute for assistant continuation.

Question 12

An incident team sees dozens of apparently completed agent runs with no explicit end signal. Which instrumentation would most directly establish whether the loop is terminating correctly?

  1. Record only how polite the final customer sentence sounds
  2. Log stop reasons, tool-use IDs and completion state
  3. Collect only the name of the selected model
  4. Count the number of comments in each generated answer

Explanation

Correct Answer

B:

A trace across request, tool execution and response enables diagnosis of missing tool returns versus genuine end_turn completion. It separates execution progress from UI assumptions.

Incorrect Answers

A:

Text quality does not establish protocol termination. A fluent sentence could coexist with outstanding calls or a truncated generation.

C:

The model name is useful metadata but cannot show how this particular run advanced through tools, errors and stopping decisions.

D:

Comment volume says little about control flow. The failure concerns loop state and stop conditions rather than writing density.

Question 13

A research assistant must gather independent facts about three regions and then write one coherent report. Which architecture best separates work and preserves synthesis accountability?

  1. Ask one worker to recursively create unlimited new workers
  2. Delegate bounded searches; have the coordinator synthesize evidence
  3. Send every search worker every other worker's full conversation
  4. Let three writers independently publish competing final reports

Explanation

Correct Answer

B:

Independent retrieval can proceed in parallel under specialist workers. A central coordinator reconciles citations and conflicting observations into the final narrative.

Incorrect Answers

A:

Unbounded delegation obscures responsibility and cost. Workload-specific fan-out limits make the research process auditable.

C:

Global transcript sharing increases context overhead and couples independent tasks. Workers generally need scoped objectives and return artifacts.

D:

Parallel final authorship does not produce the requested unified conclusion. Someone must reconcile evidence, structure and conflicts across results.

Question 14

A coordinator delegates a search and a data-analysis subtask, but both workers issue incompatible decisions about a shared calculation. Who should reconcile the conflict?

  1. Both workers by repeatedly changing each other's outputs
  2. The tool registry, because names determine which result is correct
  3. The parent coordinator, applying evidence and acceptance criteria
  4. Whichever worker finishes first, regardless of evidence
  5. Give the final decision to the worker with the highest token count

Explanation

Correct Answer

C:

The orchestrator owns integration, checks each specialist's assumptions and resolves discrepancies before finalizing. Delegation does not transfer overall correctness responsibility.

Incorrect Answers

A:

Unstructured mutual revision can oscillate and obscure provenance. A coordinator should compare evidence against an agreed criterion.

B:

A registry describes capabilities but does not evaluate conflicting empirical results. Reconciliation requires substantive analysis of returned evidence.

D:

Completion order is a scheduling property, not a measure of quality. A faster worker may have used stale or incomplete data.

E:

Token usage reflects workload rather than correctness. An integration decision needs verified evidence and explicit parent-owned acceptance criteria.

Question 15

A support team uses separate agents for refunds, account lookup and policy explanations. Simple requests are routinely sent to all three and produce duplicate backend calls. What should change?

  1. Require all agents to answer every customer message independently
  2. Route requests to a suitable specialist using a central intent decision
  3. Allow every specialist unrestricted access to the refund tool
  4. Make the models' final responses longer to reconcile differences

Explanation

Correct Answer

B:

The coordinator should select relevant workers based on the ticket requirements. Broadcasting every request wastes calls and can duplicate sensitive side effects.

Incorrect Answers

A:

Parallel complete answers still duplicate work and create disagreement. The orchestration problem is choosing the appropriate path.

C:

Wider privileges increase exposure without solving misrouting. Each worker should receive only the tools it needs for its assigned role.

D:

Longer text cannot stop the duplicate API operations that already occurred. Routing must be fixed before specialist execution.

Question 16

Four retrieval workers return overlapping reports with different document versions. What should the synthesizing coordinator prioritize?

  1. Concatenate their reports in completion order
  2. Count each worker's claim as an independent vote
  3. Discard source metadata and keep only polished prose
  4. Compare source versions and timestamps before synthesis

Explanation

Correct Answer

D:

Overlapping evidence requires normalization and conflict resolution. The coordinator should avoid treating an old observation as independent confirmation of a newer fact.

Incorrect Answers

A:

The output would preserve contradictions and duplication. Synthesis means resolving conflicts, not simply appending worker text.

B:

Workers may share the same underlying source and repeat the same error. Vote counts without source lineage inflate confidence.

C:

Removing provenance hides why conclusions differ. The coordinator needs traceable evidence to support an authoritative synthesis.

Question 17

A finance assistant has two independent read-only lookups and one funds-transfer tool that depends on both results. How should a coordinator schedule the tasks?

  1. Run the transfer alongside both lookups to minimize wall-clock time
  2. Have each lookup worker independently issue the transfer
  3. Run every task strictly one after another even when two have no dependency
  4. Run the independent lookups in parallel and gate transfer on validated results

Explanation

Correct Answer

D:

A dependency graph reduces latency without violating the prerequisite for a side effect. The money movement happens only after the reads are checked.

Incorrect Answers

A:

The action depends on information not yet established. Parallelizing it risks acting on missing or stale balances.

B:

Duplicated ownership of the side effect can trigger multiple payments. The coordinator should assign a single executor after validation.

C:

Serializing independent reads sacrifices throughput without improving the required approval of the later transfer.

Question 18

An analysis worker returns a detailed 40-page transcript, overwhelming the coordinator's context. What interface change is best?

  1. Increase every worker's context indefinitely
  2. Send the complete transcript to every sibling worker
  3. Return concise findings with evidence references and questions
  4. Remove all citations so that only conclusions remain

Explanation

Correct Answer

C:

A bounded subagent output preserves actionable knowledge without importing its entire scratchpad. Links to source artifacts allow drill-down when needed.

Incorrect Answers

A:

Larger context windows do not solve redundant handoff design. Verbose intermediate histories remain expensive and harder to reconcile.

B:

Replicating verbose histories multiplies overhead and couples tasks that can otherwise remain independent.

D:

Compression without evidence makes the synthesis uncheckable. Retain durable source pointers while reducing redundant text.

Question 19

A multi-agent document pipeline sends the writer's draft directly to customers without a final assessment. Which architecture addition most addresses quality control?

  1. Turn off evidence collection to reduce complexity
  2. Add a coordinator-owned review gate against the original requirements
  3. Give retrieval workers permission to send final customer messages
  4. Rename the writer agent to reviewer

Explanation

Correct Answer

B:

A separate evaluation stage checks correctness, formatting and evidence after delegated drafting. The parent can route revisions before publication.

Incorrect Answers

A:

Eliminating sources makes correctness harder to verify. Review depends on requirements and observations rather than fewer records.

C:

More publishing agents increase inconsistency and bypass the central synthesis responsibility.

D:

A new label does not create an independent validation pass. The same unchecked output would still be published.

Question 20

An orchestrator needs consistent handoffs among agents written by different teams. What should each worker return?

  1. The exact same prompt originally sent to the worker
  2. A return contract covering status, evidence and failures
  3. Only a subjective confidence adjective such as excellent
  4. A free-form stream of internal reasoning with no completion marker

Explanation

Correct Answer

B:

Structured contracts let the parent distinguish completed, partial and failed work. They also support predictable aggregation and retry decisions.

Incorrect Answers

A:

A prompt does not express an execution outcome. It contributes no evidence about whether the delegated task succeeded.

C:

Confidence alone does not contain findings, provenance or actionable errors. It cannot drive robust synthesis.

D:

Unbounded scratch text is difficult to parse and may omit explicit outcome status. Parent orchestration needs a reliable interface.

Claude Certified Architect - Foundations (CCAR-F) (CCA-F) Practice Tests