22 Solution Design & Architecture Practice Questions & Answers

1. 1. A support-triage service must answer within a p95 latency of 2 seconds, and its ticket categories are well defined with fixed handling steps. Which architecture fits that constraint best?

   • A.A multi-agent orchestrator that spawns a specialist worker for each ticket category and merges their competing drafts.
   • B.A single agent with an evaluator loop that critiques and rewrites each reply until a quality score passes.
   • C.A deterministic workflow that classifies the ticket, then runs that category's fixed steps with code between them.✓ Answer
   • D.An autonomous agent that plans its own steps and keeps calling tools until it decides the ticket is resolved.

   Anthropic's guidance is to find the simplest solution that works and add complexity only when it demonstrably improves outcomes. Workflows give predictable, consistent behaviour for well-defined tasks with fixed steps, whereas agentic loops trade latency and cost for task performance, and a hard p95 budget cannot absorb an unbounded number of model turns.

   Source: Anthropic, Building Effective Agents - 'Workflows vs. agents' and 'maintain simplicity'; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

2. 2. A research task spans a 20-million-token corpus that cannot fit in one context window, and its subtasks are independent of one another. What justifies an orchestrator-with-workers architecture here?

   • A.Splitting a task across agents lowers total token consumption, because no worker ever re-reads the shared context.
   • B.The work exceeds a single context window and splits into independent parallel units, so orchestration overhead is repaid.✓ Answer
   • C.Several agents cross-check one another, so any task answered by more than one model is answered more accurately.
   • D.More than one tool is involved, and a single agent cannot hold several tool definitions without confusing them.

   Anthropic's cost guidance says multi-agent orchestration repays its overhead in two situations: work that exceeds a single context window, and routine work with an expensive long tail that cheaper workers can cap. Parallel workers reading disjoint slices of a corpus is the first case; when work fits one window or forms one dependent chain, the coordination cost is not recouped.

   Source: Anthropic engineering guidance on multi-agent cost (orchestrator with parallel workers); CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

3. 3. A code-migration service receives repositories whose file counts and dependency structures vary so much that the number and nature of the subtasks cannot be known before the input is read. Which pattern fits?

   • A.Parallelisation by sectioning, where a fixed set of independent subtasks runs concurrently and is merged at the end.
   • B.Orchestrator-workers, where a central call decides the subtasks for each input and then delegates them to workers.✓ Answer
   • C.Prompt chaining, where a fixed ordered sequence of steps passes its output forward with checks placed between them.
   • D.Routing, where an initial classifier sends each incoming repository to one of several specialised fixed handlers.

   Orchestrator-workers is the named pattern for tasks whose decomposition depends on the input and cannot be enumerated in advance. Sectioning, routing and chaining all presuppose a fixed structure known before the request arrives, so none of them can express a subtask list that only exists after the repository has been inspected.

   Source: Anthropic, Building Effective Agents - orchestrator-workers pattern; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

4. 4. An invoice pipeline extracts fields from scanned PDFs and must then total the line items, apply a published tax table, and validate a check digit. How should the work be divided?

   • A.The model extracts and totals, and a cheaper model repeats the same arithmetic so the two results can be compared.
   • B.The model performs extraction and all arithmetic in one call, and a second call is asked to verify its own totals.
   • C.The model extracts the fields, and arithmetic, tax lookup and check-digit validation then run in ordinary code.✓ Answer
   • D.Regular expressions extract the fields, and the model handles the arithmetic, tax lookup and check-digit validation.

   Decomposition should give the model the step it is uniquely good at, which is reading messy scanned layouts, and leave every step with a deterministic definition in ordinary code. Arithmetic, table lookup and checksum validation have exact answers, so running them in code makes them repeatable, auditable and free of sampling variance.

   Source: Anthropic, Building Effective Agents - keep deterministic steps outside the model; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

5. 5. A marketing pipeline drafts copy, checks it against a banned-claims list, and localises it. Compliance requires that a banned claim stop the run outright. Which design meets that requirement?

   • A.A prompt chain whose middle step is a code gate that halts the run whenever a claim on the list is matched.✓ Answer
   • B.An agent holding the banned-claims list as a tool, free to decide when a check is warranted before it localises the text.
   • C.A single prompt that asks the model to draft the copy, self-check it against the list, and localise it in one response.
   • D.A voting arrangement in which three drafts are produced in parallel and the one with the fewest flagged claims is shipped.

   Prompt chaining is the pattern for decomposable sequential subtasks with programmatic gates between the steps, and a hard compliance stop is exactly such a gate. A model asked to police itself, or an agent left to decide when checking is warranted, gives a probabilistic check where the requirement is a deterministic one.

   Source: Anthropic, Building Effective Agents - prompt chaining with programmatic checks; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

6. 6. A team has two days to build an evaluation set before choosing between two models for a classification feature. Which approach yields the more decision-useful eval?

   • A.A held-out sample of clean, unambiguous inputs, so that scores are not muddied by the cases that humans also find hard.
   • B.A few dozen hand-written cases, each reviewed by two domain experts and discussed until both agree on the expected grade.
   • C.Many automated cases sampled from the real input distribution, including edge and irrelevant inputs, graded by script.✓ Answer
   • D.Cases generated by the model under test, so that their phrasing matches exactly what the deployed prompt will receive.

   Anthropic's eval guidance is counterintuitive on this point: prioritise volume over per-case polish, because many automated cases that mirror the real input distribution, edge cases included, separate two models more reliably than a small hand-graded set. Filtering out the hard or irrelevant inputs removes exactly the region where the two models differ.

   Source: Anthropic docs, Create strong empirical evaluations - prioritise volume, mirror the real distribution; CCAR-P Domain: Evaluation, Testing & OptimizationReport a problem with this question

7. 7. An agent drafts refunds, updates internal notes, sends customer emails and issues payments. The budget allows a reviewer on only some steps. Where does human review belong?

   • A.On payment issuance and outbound email, because those effects leave the system and cannot afterwards be recalled.✓ Answer
   • B.On the refund drafting step alone, because catching the first step prevents every downstream error from ever forming.
   • C.On the internal note updates, because they are the highest-volume step and errors accumulate there without being seen.
   • D.On every step, since any output can be wrong and uniform review is the easiest arrangement to explain to an auditor.

   Human-in-the-loop placement is a design decision driven by reversibility and blast radius, not a blanket fallback. Gating the irreversible, externally visible actions spends the review budget where a mistake cannot be undone, while a review queue over low-risk internal output adds cost and latency without reducing real exposure.

   Source: Anthropic guidance on human oversight of agentic actions (cost of error, reversibility); CCAR-P Domain: Governance, Safety & Risk ManagementReport a problem with this question

8. 8. A design review asks what the system does when the model API is unavailable, when it returns unparseable output, and when it declines a request. What should the architecture specify?

   • A.A circuit breaker that disables the feature on the first error and waits for an operator to re-enable it by hand.
   • B.A defined behaviour for each case: a fallback path, a bounded repair attempt, and a user-visible policy message.✓ Answer
   • C.One global retry with exponential backoff, since all three conditions are transient and clear on a later attempt.
   • D.A catch-all exception handler that logs the failure and returns one generic error so every caller sees the same result.

   These three conditions are distinct and have distinct correct responses: unavailability is an infrastructure event that a fallback or queue should absorb, an unparseable response is a formatting failure that a bounded repair step can fix, and a decline is a policy outcome the product must express to the user. Collapsing them into one retry or one generic error hides the difference and produces undefined behaviour in production.

   Source: Anthropic docs, stop_reason handling (including refusal) and SDK error classes; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

9. 9. An agent tool issues refunds. A network timeout leaves the caller unsure whether a refund was applied, and the agent loop retries the call. What keeps this safe?

   • A.Each refund attempt is logged with a timestamp so that duplicates can be found and reversed by a nightly reconciliation job.
   • B.The system prompt instructs the agent never to retry a refund, so a duplicate submission cannot arise in the first place.
   • C.The tool takes a caller-generated idempotency key, so that a repeat request returns the first result rather than refunding twice.✓ Answer
   • D.The retry uses exponential backoff with jitter, so the second attempt arrives only after the first one has certainly settled.

   Any tool with an external side effect must be safe to call more than once, because a timeout is ambiguous by nature: the request may have succeeded before the response was lost. An idempotency key supplied by the caller lets the receiving service recognise the repeat and return the original outcome, which is the only mechanism here that prevents a second refund rather than detecting or apologising for it.

   Source: Idempotent write semantics for side-effecting tools; Anthropic, Writing tools for agents - tools must be safe under retry; CCAR-P Domain: IntegrationReport a problem with this question

10. 10. A report generator makes three independent retrieval calls and then a long generation. Users say it feels slow, though none of the work can be removed. Which change helps most?

    • A.Shorten the generation and drop one of the retrievals so total wall-clock time falls under the complaint threshold.
    • B.Keep the retrievals sequential but cache each result, so that a user's second request for the report is faster.
    • C.Move the whole request onto asynchronous batch processing and email the finished report once it has completed.
    • D.Issue the three retrievals concurrently and stream the generation so that text appears as soon as it is produced.✓ Answer

    Latency architecture has two distinct levers here. Independent steps should overlap rather than queue, which cuts real wall-clock time, and streaming cuts perceived time by showing progress instead of a blank wait for the full response. Caching only helps a repeat request, and batch or trimming trade away either interactivity or the output the user asked for.

    Source: Anthropic docs, streaming for long responses; concurrency of independent steps; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

11. 11. A nightly job re-scores 400,000 archived tickets. The results are needed by morning and no user is waiting on any individual result. Which processing choice fits?

    • A.A smaller nightly sample scored interactively, with the remaining tickets left unscored in order to hold spending down.
    • B.Interactive requests fanned out across many parallel workers so that the whole run finishes as quickly as possible.
    • C.Asynchronous batch processing, which trades per-request latency for a lower cost across the same volume of work.✓ Answer
    • D.A streaming interactive path so that partial results can be watched while the overnight run is still in progress.

    Batch processing is the correct trade whenever the work is unattended and has a deadline measured in hours rather than seconds: it accepts a long, asynchronous turnaround in exchange for a materially lower price on the same tokens. Fanning out interactive calls pays the interactive rate and stresses rate limits, and sampling simply abandons part of the requirement.

    Source: Anthropic docs, Message Batches API - asynchronous processing at reduced cost; CCAR-P Domain: Evaluation, Testing & OptimizationReport a problem with this question

12. 12. After a deploy, a service's cache-read token counts drop to zero even though the wording of its system prompt did not change. The deploy added a line rendering the current time at the top of the system block. What happened?

    • A.Cache reads are only reported from the second request onward, so the counter is a reporting artefact here.
    • B.The rendered time differs on every single request, so the prefix stops matching from that point onward.✓ Answer
    • C.The cache entries expired because the deploy reset the time-to-live counter on every previously written prefix.
    • D.The system prompt fell under the minimum length that caching requires once the block had been reformatted.

    Prompt caching is a prefix match, so any byte that changes invalidates the cache from that byte onward. Placing a per-request value such as a timestamp at the top of the system block means no request ever shares a prefix with the previous one, which is the classic silent invalidator; volatile content belongs after the last breakpoint, not before it.

    Source: Anthropic docs, Prompt caching - prefix matching and cache invalidation; CCAR-P Domain: Claude Models, Prompting & Context EngineeringReport a problem with this question

13. 13. A team wants a stable cached prefix on a tool-using assistant. Which request layout holds up across requests?

    • A.The user's profile first so that each user gets a personal cache, with the shared system prompt and tools placed after it.
    • B.Tool definitions assembled per request from that user's entitlements and placed first, so the prefix matches that exact turn.
    • C.The conversation history first because it is the largest block, with the tool definitions and system prompt appended to it.
    • D.Static tool definitions, then the stable system prompt, with the cache breakpoint after the last static block.✓ Answer

    The request renders in a fixed order of tools, then system, then messages, and a change at one level invalidates that level and everything after it. Stability therefore requires the least volatile content first and the breakpoint on the last static block; a per-request tool list or a per-user block placed at the front changes the very beginning of the prefix and defeats caching for every request.

    Source: Anthropic docs, Prompt caching - render order tools, system, messages; breakpoint placement; CCAR-P Domain: Claude Models, Prompting & Context EngineeringReport a problem with this question

14. 14. A production feature meets its accuracy bar but runs well over its spending budget. Which lever should the architect evaluate before changing models?

    • A.Switching generation to a smaller model and accepting whatever accuracy the eval set then reports back.
    • B.Raising the effort setting so that answers resolve in fewer turns and the agent loop terminates earlier overall.
    • C.Caching the stable prefix and trimming the context that is resent on every single request to the model.✓ Answer
    • D.Running fewer evaluation cases in continuous integration, since that traffic is billed to the same spending line.

    Anthropic's cost work has a stated order: take the free wins before the ones that trade quality. Caching the stable prefix and cutting the tokens resent on every call reduce spend without touching output quality, so they come before effort tuning, batching, or a model change, all of which alter behaviour and require a fresh eval to justify.

    Source: Anthropic docs, Optimizing for cost and intelligence - lever order, caching first; CCAR-P Domain: Evaluation, Testing & OptimizationReport a problem with this question

15. 15. Model X costs less per token than model Y, but on the team's eval it needs more retries and more tool turns to finish a task. How should the two be compared?

    • A.By total spend per completed task on the eval set, counting the retries and the extra tool turns that were needed.✓ Answer
    • B.By average output length per call, because the model that answers more briefly is the more economical of the two.
    • C.By the published input and output prices per token, since that ratio is what determines the production bill.
    • D.By first-call success rate alone, since retries are an implementation detail of the surrounding agent loop.

    The decision-relevant unit is cost per completed task, not cost per token. A cheaper token rate that needs more turns, more retries and more re-read context to reach the same finished outcome can cost more in production, and only an eval that runs both models to completion on the same task set exposes that difference.

    Source: Anthropic docs, Optimizing for cost and intelligence - judge cost per completed task; CCAR-P Domain: Evaluation, Testing & OptimizationReport a problem with this question

16. 16. A support assistant answers from a product handbook that fits comfortably inside the model's context window and is revised twice a year. The team proposes standing up a vector database. What is the better design?

    • A.Put the whole handbook in the prompt behind a cache breakpoint and re-warm it whenever the handbook is revised.✓ Answer
    • B.Fine-tune a model on the handbook so that the knowledge is internal and no retrieval service has to be operated.
    • C.Chunk the handbook and retrieve the top passages for each query, keeping the prompt small on every single call.
    • D.Summarise the handbook into a short digest held in the system prompt, and link users to the full document for detail.

    Anthropic's explicit guidance is that a corpus small enough to sit in the context window does not need a retrieval pipeline at all: load it and use prompt caching, which removes the retrieval failure mode entirely and costs far less than an uncached read on repeat requests. Building a vector store here adds chunking, indexing and recall risk that the constraint does not require.

    Source: Anthropic, Contextual Retrieval - small knowledge bases belong in the prompt with caching; CCAR-P Domain: IntegrationReport a problem with this question

17. 17. A retrieval system answers conceptual questions well but fails when a user pastes an exact part number such as AX-4471-B. What is the likeliest cause and fix?

    • A.The embedding model is out of date, so the corpus should be re-embedded with a newer and better embedding model.
    • B.The generation model is too weak for identifier lookups, so that class of query should be routed to a stronger model.
    • C.The chunks are too small to hold the part number, so each chunk should be enlarged to cover a whole catalogue page.
    • D.Dense embeddings blur exact tokens, so lexical BM25 matching should be added and the two rankings fused together.✓ Answer

    Semantic embeddings capture meaning and therefore treat a rare alphanumeric identifier as a near-meaningless token, while lexical BM25 scoring matches it exactly. Hybrid search runs both and fuses the ranked lists, which is why it beats either method alone; this is a retrieval failure, and attributing it to the generation model is the classic misdiagnosis.

    Source: Anthropic, Contextual Retrieval - combining contextual embeddings with contextual BM25; CCAR-P Domain: IntegrationReport a problem with this question

18. 18. Chunks retrieved from a long financial filing are ambiguous on their own, because phrases like 'the quarter' and 'the segment' have no referent inside the chunk. Which change addresses this?

    • A.Retrieve more chunks for each query, so that the neighbouring context arrives alongside the ambiguous fragment.
    • B.Prepend a short generated passage that situates each chunk within its parent document before that chunk is embedded.✓ Answer
    • C.Instruct the generation prompt to ask the user a clarifying question whenever a chunk lacks a clear referent.
    • D.Lower the similarity threshold so that borderline neighbouring chunks are also admitted into the retrieved set.

    Contextual retrieval fixes ambiguity at index time rather than at query time: a short generated header situating the chunk in its parent document is prepended before embedding and lexical indexing, so the stored representation itself carries the missing referents. Widening the result set or loosening thresholds only adds more equally context-free fragments.

    Source: Anthropic, Contextual Retrieval - contextual chunk headers generated before indexing; CCAR-P Domain: IntegrationReport a problem with this question

19. 19. A long-running agent session degrades after many turns: it loses earlier decisions and repeats work it has already done. Which response matches Anthropic's context-engineering guidance?

    • A.Restart with a fresh context whenever recall degrades, and have the user restate the objective from the beginning.
    • B.Truncate the oldest turns so that the window stays small, accepting the loss of whatever those turns contained.
    • C.Move the session to a model with a larger context window so that the entire history continues to fit inside it.
    • D.Compact the history into a summary that preserves decisions and open issues, and keep notes in an external file.✓ Answer

    Recall degrades as a context grows, so attention has to be treated as a finite budget spent on signal rather than bulk. Compaction summarises the history while deliberately preserving decisions and open issues, and structured note-taking moves durable state out of the window; truncation discards that same state silently, and a larger window postpones the problem without addressing it.

    Source: Anthropic, Effective context engineering for AI agents - context rot, compaction and note-taking; CCAR-P Domain: Claude Models, Prompting & Context EngineeringReport a problem with this question

20. 20. A multi-tenant chat product resumes conversations across load-balanced stateless workers. What follows from the model API itself being stateless?

    • A.The provider retains the thread on its side, so the service needs to send only the newest user message each turn.
    • B.Sticky sessions must pin each conversation to one worker, because the history lives in that worker's own memory.
    • C.Every worker keeps a shared in-process cache of recent turns so that any worker can continue any conversation.
    • D.The service stores each conversation's history, isolated per tenant, and resends that whole history on every request.✓ Answer

    The Messages API keeps no server-side conversation state, so the full history has to be supplied on every call and therefore has to live in the application's own storage. Making that store the system of record is what lets any worker serve any turn, and keying and isolating it by tenant is what prevents one tenant's history from ever entering another tenant's prompt.

    Source: Anthropic docs, Messages API is stateless - full history sent per request; CCAR-P Domain: Solution Design & ArchitectureReport a problem with this question

21. 21. An architect wants to be able to diagnose one specific bad answer a week after it was produced. Which logging design supports that?

    • A.Per-request token usage, the prompt and model version, the retrieved chunk identifiers, and the tool-call trace.✓ Answer
    • B.A sampled one-percent capture of full request bodies, with all other traffic logged only when an exception is raised.
    • C.The final user-visible answer plus a thumbs-up or thumbs-down rating, stored against the conversation identifier.
    • D.Daily aggregates of request counts, average latency and total spend, charted per endpoint on a team dashboard.

    Diagnosing a single past answer requires the inputs that produced it, not summary statistics about traffic. Retrieval identifiers separate a retrieval miss from a prompt problem, the prompt and model version separate a regression from a deploy, and the tool-call trace shows what the model actually saw, which is why these are designed in rather than added after an incident.

    Source: Anthropic docs, usage fields on every response (input, output, cache read and creation tokens); CCAR-P Domain: Integration - observabilityReport a problem with this question

22. 22. A healthcare customer says: 'we have zero data retention enabled, so we are covered for protected health information.' How should the architect respond?

    • A.Zero data retention is sufficient once the organisation also turns off request logging on its own application servers.
    • B.HIPAA governs data held at rest only, so a service that never persists a request falls outside its scope entirely here.
    • C.Zero data retention already satisfies HIPAA, because deleting the inputs and outputs removes the protected data itself.
    • D.HIPAA readiness also requires encryption, access control and audit logging across the data lifecycle, plus a signed BAA.✓ Answer

    Zero data retention is a narrow control: inputs and outputs are not stored beyond what legal compliance and misuse enforcement require. HIPAA readiness is broader in scope, adding encryption, access control and audit logging across the whole lifecycle of protected health information, and it depends on a signed business associate agreement with a HIPAA-enabled organisation rather than on deletion alone.

    Source: Anthropic docs, API data retention and zero data retention; HIPAA compliance requires a BAA and lifecycle safeguards; CCAR-P Domain: Governance, Safety & Risk ManagementReport a problem with this question

Practice questions based on the official Claude Certified Architect – Professional (CCAR-P) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. It is delivered via Pearson VUE; Anthropic publishes the current question count, time limit, passing score and fee in the official CCAR-P exam guide. Official certification page →