CertifyAIxAce every AI certification.

CCAR-P sample questions with answers

10 free practice questions for the Claude Certified Architect: Professional exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Solution Design & Architecture

Farnell Insurance's single claims-research agent is being reviewed for a possible move to an orchestrator with subagents. The architect has gathered observations from production traces and wants only those that genuinely justify the added cost and coordination. Which TWO observations support moving to a multi-agent design? (Select TWO.)

Choose 2.

  1. A.

    Lookups return large volumes of material that is mostly irrelevant to the final judgement, and quality falls as it accumulates in context

  2. B.

    Each specialist team in the claims department wants to own a separate agent that mirrors its place in the organisation

  3. C.

    Every step of the claim assessment depends on the conclusions of the step before it and must see the full reasoning so far

  4. D.

    Each case needs searches across several independent sources whose findings do not depend on one another

  5. E.

    The team expects that splitting the work across agents will lower the total number of tokens each claim consumes

Show answer

Answer: A, D

Independent parallelisable work and context pollution from bulky lookups are the documented reasons to split.

  • A. Correct: bulky, mostly irrelevant lookups polluting context are what subagent context isolation relieves.
  • B. Incorrect: mirroring the org chart is not a technical justification and adds lossy hand-offs.
  • C. Incorrect: strict sequential dependence with shared reasoning is the documented poor fit for multi-agent designs.
  • D. Correct: independent directions of search are the parallelisable work multi-agent designs exploit.
  • E. Incorrect: multi-agent systems consume many times more tokens, not fewer.
Question 2Solution Design & Architecture

Ardley Consulting's single research agent produces briefings that meet the quality bar but take several minutes. A director asks for a multi-agent redesign, with a lead agent and parallel subagents, specifically to make briefings faster and cheaper. The briefings cover a few focused questions rather than a broad search across many sources. How should the architect respond?

  1. A.

    Proceed with the redesign, because splitting the work across parallel subagents always reduces both the elapsed time and the token cost of a task

  2. B.

    Advise against it for this goal: parallel subagents buy breadth of coverage at a substantial token cost, and on focused questions often take no less total time

  3. C.

    Proceed, but run every subagent on the smallest available model tier so that the parallel redesign is guaranteed to cost less than the single agent it would replace

  4. D.

    Advise adding subagents only for the synthesis step, since synthesis is the one stage where parallel agents reliably shorten the elapsed time

Show answer

Answer: B

Multi-agent parallelism buys thoroughness at a large token cost; it is not a general way to make focused work faster or cheaper.

  • A. Incorrect: parallel agents multiply tokens and coordination overhead, so they do not reliably cut cost or total time.
  • B. Correct: it matches the documented trade: breadth at a high token cost, with no reliable speed or cost gain on focused work.
  • C. Incorrect: a cheaper tier lowers the price per token but not the multiplied token volume, and it risks the quality bar already met.
  • D. Incorrect: synthesis integrates everyone's findings in one place; it is the least parallelisable step in the system.
Question 3Solution Design & Architecture

Brampton Health's referral-intake workflow began as a five-step chain. Over two years engineers have added hundreds of hard-coded branches for unusual referrals; each new edge case breaks an older one, and about a third of referrals now fall to manual handling because no branch fits. Common referral types still flow through the chain reliably. What migration should the architect propose?

  1. A.

    Add another layer of branches for the referrals now handled manually, so that the chain's predictable behaviour is preserved for every referral

  2. B.

    Keep the chain for the common referral types and move the unusual ones to an agent that plans its own steps with tools, tested in a sandbox with guardrails before release

  3. C.

    Replace the whole chain with an autonomous agent for every referral at once, since the growth in branching shows that fixed paths no longer work

  4. D.

    Move the branching logic out of code into the system prompt as a long list of if-then rules, so that engineers can edit it more easily

Show answer

Answer: B

Keep the workflow where the path is predictable and migrate only the unpredictable cases to an agent.

  • A. Incorrect: adding branches repeats the approach that is already breaking older cases and leaving referrals unhandled.
  • B. Correct: it keeps predictable handling where it works and gives the unpredictable cases a pattern suited to them, tested first.
  • C. Incorrect: it fixes the unusual cases but gives up predictability and adds cost on the common majority that works.
  • D. Incorrect: a prompt full of if-then rules is the brittle style context-engineering guidance warns against.
Question 4Solution Design & Architecture

Every contract summary in Garside Legal's pipeline goes to a lawyer who corrects it before release, and the team wants those corrections to improve the system over time. An engineer proposes appending each corrected summary to the prompt as a new example, which after a year would mean thousands of examples. What is the better way to feed the corrections back?

  1. A.

    Append every corrected summary as proposed, because each additional example gives the model a clearer picture of the output the lawyers expect

  2. B.

    Discard corrections once each summary is released, because lawyer review already guarantees the quality of every summary that is delivered

  3. C.

    Curate corrections into an evaluation set that gates prompt changes, and keep a small set of diverse, canonical examples in the prompt that is refreshed deliberately

  4. D.

    Append only the fifty most recent corrections, so that the prompt follows current preferences while staying within a manageable size

Show answer

Answer: C

Corrections belong in a gating evaluation set and a curated example set, not an ever-growing prompt.

  • A. Incorrect: thousands of appended examples bloat the context, contrary to the advice to curate canonical examples.
  • B. Incorrect: it throws away the best available signal and leaves reviewers fixing the same errors forever.
  • C. Correct: corrections both gate future changes as tests and inform a small, deliberately curated example set.
  • D. Incorrect: a recency window bounds size but is not curation, drops rare important patterns and ignores testing.
Question 5Solution Design & Architecture

Stanway Marketplace screens new listings with three independent Claude prompts that vote on whether a listing is prohibited. The same check serves two categories. In children's toys, a prohibited item reaching buyers is a serious safety incident. In second-hand books, prohibited items are rare and a wrongly blocked listing is the main harm, costing sellers sales. How should the vote thresholds be set?

  1. A.

    Hold a listing in either category when any one prompt flags it, because the value of voting is catching whatever a single prompt notices

  2. B.

    Hold a toy listing when any one prompt flags it, and require agreement from more of the prompts before holding a book listing

  3. C.

    Use the same majority threshold in both categories, so that sellers in each category are treated by exactly the same rule

  4. D.

    Require a majority for toy listings and hold book listings when any one prompt flags them, since books generate most seller complaints

Show answer

Answer: B

Voting thresholds are a dial between false positives and false negatives; set each by the cost of each error.

  • A. Incorrect: any-flag holding minimises misses but over-blocks books, where wrongful holds are the main harm.
  • B. Correct: each category's threshold follows its costlier error: misses in toys, wrongful blocks in books.
  • C. Incorrect: consistent-looking, but one threshold cannot suit two categories whose error costs point in opposite directions.
  • D. Incorrect: it inverts the logic, relaxing the safety-critical category and tightening the one where blocks cause the harm.
Question 6Solution Design & Architecture

Anselm Travel supports customers in nine languages and will add four more next year. Its intent classifier is a separately trained model per language, each needing its own labelled data and retraining whenever intents change. The platform team is two engineers. Leadership wants lower maintenance cost without any loss of coverage or accuracy in existing languages. Which plan best meets this?

  1. A.

    Keep the per-language models and contract a labelling vendor for each new language, so every classifier retains its present accuracy

  2. B.

    Keep the per-language models, and translate tickets in the four new languages into an existing language first so that no new models need training

  3. C.

    Replace all nine models at once with one Claude classifier, because a single multilingual model removes any need to evaluate the languages separately

  4. D.

    Consolidate on one Claude-based classifier with shared intent definitions, and evaluate accuracy per language on labelled samples before retiring each old model

Show answer

Answer: D

One multilingual classifier cuts the per-language maintenance burden, but each language must be measured before its old model is retired.

  • A. Incorrect: per-language labelling and retraining is exactly the recurring cost leadership wants to remove.
  • B. Incorrect: a translation stage is another component to maintain and a new source of error, not a reduction in upkeep.
  • C. Incorrect: it consolidates, but skipping per-language evaluation risks unseen accuracy loss that the constraint forbids.
  • D. Correct: it cuts maintenance to one classifier while measuring each language before retirement, so accuracy is protected.
Question 7Solution Design & Architecture

Penhallow Insurance routes claims correspondence with a rules engine that matches policy numbers and form codes. It routes most items correctly, but free-text letters with no form code fall into a manual queue that keeps growing. Leadership asks the architect to replace the rules with AI. The rules engine is well tested and cheap to run. Which design best translates this into a first release?

  1. A.

    Keep the rules engine and add hand-written rules for the free-text letters until the manual queue returns to its earlier size

  2. B.

    Send every item to both the rules engine and Claude and route each one by whichever answer arrives first, so throughput stays the same

  3. C.

    Replace the rules engine with Claude for all correspondence, so that a single system owns routing and the tested rules no longer need maintaining by anyone

  4. D.

    Keep the rules engine on the traffic it routes correctly and send only the items it cannot route to a Claude classifier, measured against the manual queue's outcomes

Show answer

Answer: D

Target Claude at the cases the existing rules fail on, and leave the tested, cheap rules where they already work.

  • A. Incorrect: free-text letters are the input that pattern-based rules handle worst, so more rules will not drain the queue.
  • B. Incorrect: choosing the fastest answer makes the route depend on latency rather than on which system is right.
  • C. Incorrect: workable, but it pays to re-solve traffic that is already routed correctly, adding cost and risk the stated constraint argues against.
  • D. Correct: it applies Claude only where the rules fail and keeps a proven, cheap component on the traffic it already handles.
Question 8Solution Design & Architecture

Colbrook Software runs a voting check, several independent Claude reviews of each change, to catch security vulnerabilities before merge. Most changes touch only documentation or tests, and the check's cost has become a line item leadership questions. Missing a vulnerability in authentication or payment code remains unacceptable. Which redesign best balances these constraints?

  1. A.

    Classify each change first, send changes touching security-sensitive code through the full voting check, and give low-risk changes a single review

  2. B.

    Run the voting check on a random sample of changes each week, and treat the sample's findings as representative of every change

  3. C.

    Keep full voting on every change but move every review to the smallest model tier, so cost falls without any change to the design

  4. D.

    Reduce the check to a single review for every change, since most changes are low risk and one review usually agrees with the majority anyway

Show answer

Answer: A

Combine patterns: route by risk so that expensive voting is spent only where a miss is unacceptable.

  • A. Correct: routing concentrates the costly voting on the changes where a missed vulnerability is unacceptable.
  • B. Incorrect: sampling leaves most sensitive changes without the check, contradicting the no-miss requirement.
  • C. Incorrect: cheaper, but it weakens the review exactly where misses are unacceptable and still votes on trivial changes.
  • D. Incorrect: it drops the extra confidence on the security-sensitive changes where it is required.
Question 9Solution Design & Architecture

Tarrant Logistics lets Claude read failed evaluation transcripts and rewrite its agent's tool descriptions, then re-runs the same evaluation tasks to confirm each improvement. After several rounds the score on those tasks is excellent, but task completion in production has barely moved. Which change to this improvement loop matters most?

  1. A.

    Run further rounds of rewriting on the same tasks until the score stops rising, since the gains will then start to appear in production

  2. B.

    Measure each round's rewrites on a held-out set of realistic tasks the rewriting step never sees, and accept only changes that improve it

  3. C.

    Replace the evaluation tasks with simple sandbox tasks that isolate one tool each, so every rewrite can be judged on a single tool

  4. D.

    Stop using Claude to rewrite the descriptions, because a model editing its own tools cannot produce changes that generalise beyond what it saw

Show answer

Answer: B

A self-improvement loop needs a held-out set, or it optimises for the tasks it was shown.

  • A. Incorrect: more rounds on the same tasks deepen the overfitting that is already visible.
  • B. Correct: a held-out set of realistic tasks shows whether rewrites generalise instead of fitting the tasks they were tuned on.
  • C. Incorrect: the guidance warns that simple sandbox tasks fail to stress-test tools with realistic complexity.
  • D. Incorrect: Anthropic endorses Claude-driven tool improvement; the flaw is how improvement is measured, not who rewrites.
Question 10Solution Design & Architecture

Corrin Bank has shortlisted four candidate processes for its first Claude agent. The sponsors want a release that can prove its value within one quarter, and it must be defensible to the bank's risk committee, which reviews any automated decision affecting customers. Which candidate is the strongest first release?

  1. A.

    Resolving card-dispute enquiries, where each case ends in a recorded resolution, the agent works through existing case tools, and staff review escalations

  2. B.

    Approving small-business loans end to end without staff involvement, which would remove more manual effort than any other process on the list

  3. C.

    Drafting the bank's annual strategy narrative, which executives regard as the highest-value writing task in the organisation and would welcome help with

  4. D.

    Summarising internal meeting notes for all staff, which reaches the most employees and needs no integration with any core banking system

Show answer

Answer: A

Pick a process with measurable resolutions, tool-based actions and human oversight, the profile Anthropic describes for agents that deliver value.

  • A. Correct: measurable resolutions, action through tools and staff review of escalations satisfy both the value and the risk constraint.
  • B. Incorrect: it removes human oversight from a consequential decision, which the risk committee constraint rules out.
  • C. Incorrect: quality is judged by taste, so there is no clear success criterion that could prove value in a quarter.
  • D. Incorrect: low risk, but it offers no clear success measure, so it cannot demonstrate value on the sponsors' terms.

Keep going with 490 more CCAR-P questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all CCAR-P papers
CCAR-P sample questions with answers (10 free) · CertifyAIx