16 Solution Design & Architecture Practice Questions & Answers

1. 1. A retail bank must produce a monthly regulatory compliance summary: extract figures from a fixed set of reports, normalize them, draft the narrative, and verify the numbers against the sources. The steps are identical every month and each intermediate output can be checked programmatically. Which architecture should the senior architect recommend?

   • A.A single LLM call containing the entire task in one long prompt
   • B.A fully autonomous agent that decides at run time which reports to read and when to stop
   • C.An orchestrator-workers system that dynamically re-decomposes the reporting task each month
   • D.A prompt-chaining workflow with programmatic gates that validate each intermediate output✓ Answer

   When a task decomposes cleanly into fixed, predictable subtasks, Anthropic's guidance is to use a workflow with predefined code paths rather than agentic autonomy: prompt chaining makes each LLM call an easier task (trading latency for accuracy) and programmatic gates catch errors between steps. An agent or dynamic orchestrator adds cost, latency, and unpredictability with no benefit when the steps never vary, and one giant prompt forgoes the accuracy gains of decomposition.

   Source: Anthropic, "Building Effective Agents" — Workflow: Prompt chaining; "find the simplest solution possible" principleReport a problem with this question

2. 2. A hospital patient portal receives messages that fall into clearly distinct categories — appointment scheduling, billing disputes, and clinical questions — each needing different prompts, tools, and escalation rules. A single combined prompt has degraded quality across all three categories. Which architectural pattern best fits?

   • A.An evaluator-optimizer loop that iteratively refines one general-purpose response
   • B.Parallelization with voting across three identical model instances
   • C.A fully autonomous agent that freely chooses tools for every message
   • D.Routing: classify each incoming message, then direct it to a specialized downstream flow for its category✓ Answer

   Routing is designed for exactly this situation: it classifies an input and directs it to a specialized follow-up task, providing separation of concerns so each category gets its own optimized prompt and tools. The underlying mechanism is that optimizing one input type in a shared prompt hurts performance on the others; routing works well when categories are distinct, are handled better separately, and can be classified accurately.

   Source: Anthropic, "Building Effective Agents" — Workflow: Routing (separation of concerns; distinct categories handled better separately)Report a problem with this question

3. 3. A national retailer moderates user-submitted product reviews. Legal requires very high confidence before a review is flagged, and the team can afford several model calls per review. Which parallelization approach directly addresses the confidence requirement?

   • A.Routing each review to a category-specific prompt before a single moderation call
   • B.Voting: run the moderation task multiple times with prompts evaluating different aspects, and aggregate the verdicts under a decision threshold✓ Answer
   • C.Prompt chaining where one moderation call feeds its verdict into the next
   • D.An orchestrator that spawns workers to rewrite the offending reviews

   The voting variant of parallelization runs the same task multiple times to obtain diverse perspectives and aggregates the results, which is Anthropic's cited approach for content-flagging: multiple prompts evaluate different aspects and a vote threshold balances false positives against false negatives. Chaining and routing do not produce independent judgments to aggregate, and rewriting content does not address the confidence requirement at all.

   Source: Anthropic, "Building Effective Agents" — Workflow: Parallelization, voting variant (flagging content with multiple prompts and vote thresholds)Report a problem with this question

4. 4. An engineering-services firm wants Claude to implement code changes from tickets, where each ticket may require modifying an unpredictable number and combination of files. Why is the orchestrator-workers pattern more appropriate here than fixed parallelization?

   • A.Because the subtasks are not known in advance — a central LLM must determine them from each ticket, delegate them to workers, and synthesize the results✓ Answer
   • B.Because parallelized model calls are unable to invoke tools
   • C.Because orchestrator-workers always runs faster and cheaper than parallelization
   • D.Because voting across multiple workers guarantees the code is correct

   Anthropic identifies the key difference explicitly: in orchestrator-workers the subtasks are not pre-defined but are determined dynamically by the orchestrator based on the specific input, which is why coding tasks that touch an unpredictable set of files are the canonical example. Parallelization requires subtasks that can be fixed in advance; voting does not guarantee correctness, parallel calls can use tools, and the orchestrator pattern typically costs more, not less.

   Source: Anthropic, "Building Effective Agents" — Workflow: Orchestrator-workers (subtasks determined dynamically by the orchestrator, coding example)Report a problem with this question

5. 5. A financial-services firm generates client-facing market commentary that must satisfy a documented style and accuracy checklist, and drafts measurably improve when they are critiqued and revised. Which pattern best exploits these conditions?

   • A.A single call with chain-of-thought instructions and no review step
   • B.Routing each commentary request to a category-specific prompt
   • C.An autonomous agent that browses market data sources indefinitely
   • D.Evaluator-optimizer: one LLM generates the draft while another evaluates it against the checklist, looping until the criteria are met✓ Answer

   The evaluator-optimizer workflow is effective precisely when two conditions hold: clear evaluation criteria exist (the documented checklist) and iterative refinement provides measurable value (drafts improve when critiqued). Pairing a generator LLM with an evaluator LLM in a loop mirrors a human writer's revision cycle; routing addresses input diversity, not quality refinement, and an unbounded agent adds autonomy the task does not need.

   Source: Anthropic, "Building Effective Agents" — Workflow: Evaluator-optimizer (use when clear evaluation criteria exist and iterative refinement provides measurable value)Report a problem with this question

6. 6. At project kickoff, an insurance company's engineering team proposes standing up a multi-agent framework before requirements are fully understood. According to Anthropic's architectural guidance, what should the senior architect recommend first?

   • A.Fine-tune a custom model before writing any prompts
   • B.Adopt the multi-agent framework immediately to future-proof the platform
   • C.Start with a single augmented LLM call — the model enhanced with retrieval, tools, and memory — and add multi-step orchestration only if measured results fall short✓ Answer
   • D.Build an orchestrator-workers topology from day one so it never needs re-architecting

   Anthropic's core guidance is to find the simplest solution possible and only increase complexity when it demonstrably improves outcomes: the augmented LLM (an LLM enhanced with retrieval, tools, and memory) is the basic building block, and many production successes use simple, composable patterns rather than frameworks. Committing to multi-agent infrastructure before requirements and measurements exist optimizes for sophistication instead of fit, adding cost and latency without proven benefit.

   Source: Anthropic, "Building Effective Agents" — The augmented LLM as basic building block; "find the simplest solution possible, and only increase complexity when needed"Report a problem with this question

7. 7. A government agency is designing automation for its internal IT service desk. Troubleshooting requests are open-ended, the number of diagnostic steps cannot be predicted in advance, resolution can be verified against system state, and staff can approve risky actions. Which factor most strongly justifies an agentic design here?

   • A.The number of steps cannot be predicted and a fixed path cannot be hardcoded, while the agent can check progress against environment feedback at each step✓ Answer
   • B.Agents cost less per request than predefined workflows
   • C.Agents remove the need for any human oversight of risky actions
   • D.Tool use by agents eliminates the possibility of hallucination

   Anthropic's criteria for choosing agents are open-ended problems where the number of steps is difficult or impossible to predict and a fixed path cannot be hardcoded — exactly this scenario — combined with the agent's ability to gain ground truth from the environment (tool results, system state) to assess its own progress. The distractors invert the actual trade-offs: agents typically cost more than workflows, still require guardrails and human checkpoints, and tool use does not eliminate hallucination.

   Source: Anthropic, "Building Effective Agents" — Agents section (open-ended problems, unpredictable number of steps, ground truth from the environment)Report a problem with this question

8. 8. A healthcare payer plans to deploy an autonomous agent that adjudicates claims, including actions that are difficult to reverse once executed. Per Anthropic's guidance on autonomous agents, what must the architecture include before production rollout?

   • A.As many tools as possible so the agent never gets stuck mid-task
   • B.A higher temperature setting so the agent handles novel claims more creatively
   • C.Extensive testing in sandboxed environments plus guardrails, with checkpoints for human review before high-impact, hard-to-reverse actions✓ Answer
   • D.Removal of human checkpoints to maximize straight-through processing

   Anthropic warns that the autonomous nature of agents means higher costs and the potential for compounding errors, and therefore recommends extensive testing in sandboxed environments along with appropriate guardrails; agents can also pause for human feedback at checkpoints. Because claim adjudication includes hard-to-reverse actions, the correct control is validating behavior safely and gating irreversible steps — not increasing randomness, expanding the tool surface, or removing oversight.

   Source: Anthropic, "Building Effective Agents" — Agents section (sandboxed testing, guardrails, checkpoints for human feedback; compounding-error risk)Report a problem with this question

9. 9. A retailer has deployed a Claude-based product-description generator: the input pipeline, the processing stage, and output publishing all work. Under the end-to-end architecture model (input → processing → output → feedback), which addition completes the design?

   • A.Capturing downstream outcomes — editor corrections, engagement, and return rates — and feeding them into evaluation datasets that drive prompt and architecture iteration✓ Answer
   • B.A larger context window so longer product data fits in one call
   • C.Prompt caching on the static prefix to reduce cost
   • D.A second model running in parallel purely for redundancy

   The feedback stage is what makes the architecture a closed loop: real-world outcome signals flow back into evaluation and improvement, allowing the team to measure performance, detect drift or regressions, and iterate on prompts and patterns — Anthropic's guidance is to measure performance and iterate on implementations. Redundancy, caching, and context size are point optimizations of processing; none of them returns outcome information to the system, so none completes the input → processing → output → feedback cycle.

   Source: CCAR-P Exam Guide, Domain 1 — "Design end-to-end architectures (input → processing → output → feedback loops)"; Anthropic, "Building Effective Agents" (measure performance and iterate)Report a problem with this question

10. 10. A retail chatbot must meet a strict, contractually defined response-latency SLA. Two candidate designs both meet the quality bar in evaluation: (1) a single augmented-LLM call with retrieval, and (2) a multi-step agent that averages many sequential model calls per response. Which design should the architect select, and why?

    • A.The single augmented-LLM call — agentic multi-step designs trade latency and cost for capability, and here the extra sequential calls consume the latency budget without needed quality gains✓ Answer
    • B.Either design, because latency is unaffected by the number of sequential model calls
    • C.The multi-step agent, because adding steps always increases answer accuracy
    • D.The multi-step agent, because SLAs concern only service uptime, not response time

    Anthropic states that agentic systems often trade latency and cost for better task performance, and advises choosing the simplest design that meets requirements; when both designs clear the quality bar, the performance-SLA pillar makes latency the deciding constraint, and each additional sequential model call adds latency. Selecting the agent would spend latency budget on capability the evaluation shows is not needed, violating both the SLA alignment and the simplicity principle.

    Source: Anthropic, "Building Effective Agents" — "agentic systems often trade latency and cost for better task performance"; CCAR-P Exam Guide, Domain 1 (performance SLAs value pillar)Report a problem with this question

11. 11. A retailer's marketing team uses Claude to produce localized campaign copy. A single prompt that drafts the copy and translates it in one shot yields inconsistent brand voice. Which decomposition correctly applies the prompt-chaining technique?

    • A.Send the same combined prompt to several models and keep the longest output
    • B.Let an autonomous agent decide on each request whether to translate before or after drafting
    • C.Keep everything in one call but instruct the model to think step by step
    • D.Split the work into sequential subtasks — draft the copy, run a programmatic gate that checks it against brand criteria, then translate — accepting extra latency because each easier call is more accurate✓ Answer

    Prompt chaining decomposes a task into a sequence of steps where each LLM call processes the output of the previous one, optionally with programmatic checks ('gates') on intermediate outputs to keep the process on track; Anthropic explicitly frames it as trading latency for higher accuracy by making each LLM call an easier task, and cites writing-then-translating as an example. Chain-of-thought inside one call does not create checkable intermediate outputs, and the other options add no gated sequential structure at all.

    Source: Anthropic, "Building Effective Agents" — Workflow: Prompt chaining (gates on intermediate steps; trading latency for accuracy; write-then-translate example)Report a problem with this question

12. 12. A government agency tells its consulting architect only: "We want AI to modernize our permitting process." Before selecting any model or architectural pattern, what should the architect do first?

    • A.Build a multi-agent prototype quickly to impress stakeholders and win buy-in
    • B.Define measurable business outcomes and success criteria (e.g., cycle-time and error-rate targets) and identify which steps of the permitting process are actually suited to an LLM✓ Answer
    • C.Choose the integration protocol and connectors before anything else
    • D.Select the most capable available model so it can absorb whatever requirements emerge

    Translating a business problem into a Claude solution starts with the problem, not the technology: success criteria and measurable outcomes determine which pattern is 'right,' because Anthropic's guidance is that success is not building the most sophisticated system but the right system for your needs, with performance measured and iterated against those criteria. Choosing models, protocols, or a flashy multi-agent prototype before the outcome is defined optimizes unknowns and risks solving the wrong problem.

    Source: CCAR-P Exam Guide, Domain 1 — "Translate business problems into Claude-based AI solutions"; Anthropic, "Building Effective Agents" — "success ... isn't about building the most sophisticated system. It's about building the right system for your needs"Report a problem with this question

13. 13. A healthcare benefits assistant must both answer members' coverage questions and screen every exchange for protected health information (PHI) leakage and inappropriate content. Combining both duties in one prompt has measurably lowered screening recall. Which pattern addresses this?

    • A.An evaluator-optimizer loop that rewrites each answer until it is polite
    • B.Parallelization by sectioning: one model instance answers the coverage question while a separate instance screens the content, so each call handles a single focused concern✓ Answer
    • C.One larger prompt with more detailed screening instructions appended at the end
    • D.Routing members to different queues based on their plan type

    Anthropic's sectioning variant of parallelization cites exactly this use case: implementing guardrails where one model instance processes user queries while another screens them, because separate calls handling each consideration perform better than one call responsible for everything — LLMs attend better when each concern is focused. An evaluator loop targets output quality, routing targets input categories, and piling more instructions onto the shared prompt is the very design that caused the recall drop.

    Source: Anthropic, "Building Effective Agents" — Workflow: Parallelization, sectioning variant (guardrails example: one instance processes queries while another screens them)Report a problem with this question

14. 14. A financial-services support system receives a mix of simple refund-policy questions and deep technical troubleshooting requests. Every attempt to tune the shared prompt for one traffic type degrades quality on the other, and a lightweight classifier distinguishes the two reliably. Under what condition does Anthropic say routing is the right choice — and is it met here?

    • A.Routing fits when inputs fall into distinct categories that are handled better separately and classification is accurate — both conditions hold here, so routing prevents optimizing one category at the expense of the others✓ Answer
    • B.Routing fits only when all inputs are homogeneous — so it does not apply here
    • C.Routing fits only when inputs cannot be classified accurately — so it does not apply here
    • D.Routing fits only when no classification step can be afforded — so it does not apply here

    Anthropic states that routing works well for complex tasks where there are distinct categories that are better handled separately, and where classification can be handled accurately by an LLM or a traditional model; the mechanism is separation of concerns, letting each category have specialized prompts (and even differently sized models) so that optimizing for one input type no longer degrades performance on the others. The scenario exhibits both conditions, so routing is directly indicated.

    Source: Anthropic, "Building Effective Agents" — Workflow: Routing ("works well for complex tasks where there are distinct categories that are better handled separately, and where classification can be handled accurately")Report a problem with this question

15. 15. A document-processing agent built for a government archive plans its entire multi-step task upfront, then executes every step without examining any results until the end; on long tasks it drifts badly from the goal. Which redesign aligns with Anthropic's guidance on how agents should operate?

    • A.Remove the agent's tools so it relies solely on its internal knowledge
    • B.Have the agent gain ground truth from the environment at each step — tool-call results and validation output — to assess its progress, and pause at checkpoints for human feedback on long runs✓ Answer
    • C.Produce an even longer, more detailed upfront plan so nothing is left ambiguous
    • D.Raise the maximum output tokens so the full plan always fits in one response

    Anthropic describes agents as operating in a loop: during execution it is crucial that the agent gains 'ground truth' from the environment at each step — such as tool call results or code execution — to assess its progress, and agents can pause at checkpoints or when hitting blockers for human feedback. Pure upfront planning cannot correct for surprises the environment reveals mid-task, which is exactly why the current design drifts; removing tools or enlarging output limits does nothing to restore per-step feedback.

    Source: Anthropic, "Building Effective Agents" — Agents section (gain ground truth from the environment at each step; pause at checkpoints for human feedback)Report a problem with this question

16. 16. A CIO asks the architect to position the business case for a Claude solution that automates the company's existing invoice-triage process, cutting cycle time and manual touches without changing what the business offers to customers. Under the business value pillars, where does this case primarily belong?

    • A.Transformation — any adoption of AI is by definition transformative
    • B.Cost — the case rests primarily on reducing model token spend
    • C.Efficiency — it streamlines an existing process, reducing cycle time and manual effort, without creating a new business capability✓ Answer
    • D.Performance SLAs — the case rests on new contractual latency commitments

    The efficiency pillar covers making existing processes faster and less labor-intensive, which is exactly what streamlining invoice triage does; transformation applies when AI enables a genuinely new capability or business model, which this case explicitly does not. Mapping the initiative to the correct pillar matters architecturally because it fixes the success metrics (cycle time, manual-touch rate) that the design and its feedback loops must be built to measure.

    Source: CCAR-P Exam Guide, Domain 1 — "Align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs)"Report 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. The real exam is 63 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($175). Official certification page →