Part 9a — Question Bank: Domain 1, Solution Design & Architecture (Q1–Q26)
Original practice questions. Not official, not leaked, not from the live item bank. Official objectives are quoted from Exam Guide v1.0 §6. Format note: The real exam mixes single-answer and multiple-response items, and each item states how many to select. Single-answer items here have 4 choices (A–D). Multiple-response items here have 5 choices (A–E) and say how many to select; the 5-choice format is my choice for practice, since the guide doesn't specify option counts.
Q1 · D1 · Objective: Translate business problems into Claude-based AI solutions · Difficulty: Intermediate
Scenario: A retailer's COO asks for "AI to cut returns fraud." The team has years of labeled structured transaction data, needs a decision in under 50 ms at checkout, and also has free-text return notes that analysts rarely read. Question: What is the best first move? A. Build an autonomous Claude agent that reviews every return and blocks suspicious ones. B. Assess task fit: use conventional scoring/rules for the structured, latency-critical decision, and evaluate Claude only for the unstructured return-note signal. C. Send each customer's full transaction history to the largest model on every return. D. Delay the project until models offer lower latency than 50 ms. Correct answer: B Explanation: The core decision is structured, labeled, and latency-bound — a classic fit for conventional ML/rules. Claude adds value on unstructured text where variability is high. Why not the others: A hands an irreversible, latency-critical decision to an autonomous LLM with no case for it. C is slow and costly and sends excess data. D assumes Claude must own the entire flow. Architectural principle: Use LLMs where input variability is high; keep deterministic or classic-ML paths where they fit better. Ref: Prep course Lesson 1; https://www.anthropic.com/research/building-effective-agents
Q2 · D1 · Objective: Align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs) · Difficulty: Intermediate
Scenario: A hospital network has no overnight patient-intake service today. It wants Claude to offer 24/7 intake in six languages. Question: Which framing and metric best match the value being created? A. Cost: dollars saved per intake call. B. Productivity: nurse hours saved per week. C. Transformation: a capability that did not exist — measure coverage (hours/languages served), completion rate, and escalation quality. D. Performance SLA: average response time of the chat widget. Correct answer: C Explanation: The primary value is a new capability, so the success metrics should reflect coverage and outcomes of that new service rather than savings against a baseline that doesn't exist. Why not the others: A and B assume a baseline process being made cheaper or faster. D is a supporting NFR, not the value case. Architectural principle: Pick the value pillar that matches the sponsor's real objective, then choose metrics to match. Ref: Exam Guide §6 D1; Prep Lesson 4
Q3 · D1 · Objective: Translate business problems into Claude-based AI solutions · Difficulty: Advanced
Scenario: Legal operations asks to "automate contract review." Attorneys must remain accountable; there is a clause playbook for NDAs only; a labeled sample of 200 NDAs exists. Question: Which scope is the most defensible first release? A. Flag and extract playbook-defined clause deviations in NDAs with attorney review, measured on recall against the labeled set. B. Negotiate and redline all contract types autonomously. C. Summarize every contract in the repository into one paragraph. D. Replace attorney review for contracts under a value threshold. Correct answer: A Explanation: It's a thin slice with data, a playbook, a measurable target, and human accountability. Why not the others: B and D remove accountability and exceed the available playbook/data. C is measurable on nothing and creates no decision value. Architectural principle: Scope to the thinnest slice with ground truth, measurable success, and human oversight. Ref: Prep Lesson 1; https://platform.claude.com/docs/en/test-and-evaluate/define-success
Q4 · D1 · Objective: Translate business problems into Claude-based AI solutions · Difficulty: Advanced
Scenario: A sponsor wants contract-risk analysis of 100-page documents to be "instant, perfectly accurate, and cheap." Question: What should the architect do? A. Use the most capable model for every request and promise the sponsor all three goals. B. Quietly prioritize accuracy and let latency slip. C. Shorten documents until all three goals look attainable. D. Surface the conflict as explicit trade-offs and propose tiers — e.g., fast low-cost triage, plus deeper asynchronous analysis with review — quantified with measured numbers. Correct answer: D Explanation: Conflicting NFRs must be made visible and resolved by the stakeholder, with options that map to levers (model tier, caching, async). Why not the others: A over-promises. B makes a hidden decision for the business. C silently changes scope. Architectural principle: Expose competing non-functional requirements and let the business choose among priced options. Ref: Prep Lesson 4; Exam Guide §6 D6
Q5 · D1 · Objective: Translate business problems into Claude-based AI solutions · Difficulty: Intermediate
Scenario: A team built a demo and asks which model to standardize on. There are no acceptance criteria and no test set. Question: What should happen before choosing a model? A. Choose the newest top-tier model to be safe. B. Define measurable acceptance criteria and build a representative eval set, then compare models against them. C. Ask users which model name they prefer. D. Pick the cheapest model and adjust prompts until the demo looks good. Correct answer: B Explanation: Model selection is an empirical decision against criteria; without criteria there's no basis for a defensible choice. Why not the others: A and D substitute assumptions for evidence. C is irrelevant to measurable quality. Architectural principle: Evals as acceptance criteria precede model and architecture decisions. Ref: https://platform.claude.com/docs/en/test-and-evaluate/develop-tests
Q6 · D1 · Objective: Translate business problems into Claude-based AI solutions · Difficulty: Advanced
Scenario: A tax-software vendor must produce identical results for identical inputs; calculations follow fixed published formulas; customers upload messy PDFs of source documents. Question: Which design is most appropriate? A. Have Claude compute the tax figures directly from the PDFs. B. Have Claude compute figures twice and accept them when both runs agree. C. Use Claude to extract inputs from documents with validation, and compute tax in deterministic code; optionally use Claude to explain results. D. Fine-tune a model on past returns so it learns the formulas. Correct answer: C Explanation: Extraction from variable inputs suits an LLM; fixed formulas and reproducibility call for deterministic code. Why not the others: A and B put arithmetic and rules in a probabilistic component; agreement isn't correctness. D is an unnecessary path that still isn't deterministic. Architectural principle: Keep deterministic logic deterministic; use LLMs for language-heavy, variable tasks. Ref: Building effective agents; Prep Lesson 1
Q7 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Scenario: A nightly process extracts eight fields from invoices that arrive in three known layouts. Nobody waits on the result. Question: Which pattern is best? A. Route by layout to a layout-specific extraction prompt, run as a non-interactive batch with schema validation. B. A fully autonomous agent that decides how to process each invoice. C. A single prompt handling all layouts, run synchronously per invoice with the largest model. D. A multi-agent debate among several extraction agents per invoice. Correct answer: A Explanation: Known categories + no latency pressure = routing/workflow with batch for cost efficiency. Why not the others: B and D add cost and variance without need; C ignores layout-specific tuning and the cost advantage of batch. Architectural principle: Use the simplest pattern that meets the requirements; match processing mode to the SLA. Ref: Building effective agents; https://platform.claude.com/docs/build-with-claude/batch-processing
Q8 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Advanced
Scenario: An SRE team wants help diagnosing production anomalies. Investigation paths are unpredictable; every hypothesis can be checked by read-only queries to logs and metrics. Question: Which design is justified? A. A fixed five-step prompt chain that always checks the same dashboards. B. A single prompt with the last hour of logs pasted in. C. A router sending each alert to one of three canned runbooks. D. An agent with read-only diagnostic tools, step/token/time budgets, verifiable progress checks, and human escalation. Correct answer: D Explanation: The path can't be predefined but progress is verifiable, which is where agents fit — with budgets and read-only tools to bound risk. Why not the others: A and C assume known paths. B can't gather evidence on demand. Architectural principle: Use agents when flexibility is needed and progress is verifiable; bound them with budgets and least privilege. Ref: Building effective agents
Q9 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Scenario: A marketing team needs product descriptions that must satisfy a written brand rubric (tone, banned claims, length). Reviewers currently send drafts back several times. Question: Which workflow pattern fits best? A. Routing by product category. B. Evaluator-optimizer: generate, grade against the rubric, revise, with an iteration cap. C. Parallelization to produce ten drafts and pick randomly. D. An open-ended agent with web access. Correct answer: B Explanation: Clear evaluation criteria and iterative refinement are the evaluator-optimizer sweet spot. Why not the others: A doesn't address quality iteration. C lacks a quality criterion for selection. D adds autonomy and risk with no benefit. Architectural principle: Use iterative refinement when criteria are explicit and revision demonstrably helps. Ref: Building effective agents
Q10 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Scenario: Support tickets fall into five categories; each needs different instructions, tools, and policy snippets. A single mega-prompt has become unwieldy and error-prone. Question: What is the best restructuring? A. Add more rules to the mega-prompt. B. Give the model all tools in every case. C. Add a routing step that classifies the ticket and dispatches to a specialized prompt/toolset per category. D. Convert to a multi-agent swarm that negotiates category membership. Correct answer: C Explanation: Routing separates concerns, allows per-category prompts and least-privilege tools, and is easy to evaluate. Why not the others: A worsens prompt complexity. B widens capability bloat. D adds coordination overhead. Architectural principle: Separate concerns with routing when inputs fall into distinct categories. Ref: Building effective agents
Q11 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Scenario: A reviewer must check a document for four independent risks (privacy, export control, IP, safety claims). Latency matters, and high-stakes misses are costly. Question: Which design is most appropriate? A. Parallel independent checks (one focused prompt per risk), with an aggregation step. B. One long sequential chain where each step re-reads previous outputs. C. An agent that decides which risks to look at. D. Skip decomposition and ask one prompt to "find all risks." Correct answer: A Explanation: Independent subtasks can run in parallel, with focused prompts improving reliability and reducing wall-clock time. Why not the others: B adds needless latency and coupling. C introduces autonomy where the risk list is known. D reduces focus and recall. Architectural principle: Parallelize independent subtasks; keep prompts focused. Ref: Building effective agents
Q12 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Advanced
Scenario: A research agent's costs vary wildly; some runs loop for 90 tool calls repeating similar searches.
Question: Which change best addresses the problem?
A. Increase max_tokens.
B. Switch to the largest model so it "reasons better."
C. Remove tool descriptions to save tokens.
D. Add step, token, and time budgets with explicit stopping conditions and progress checks; consider a workflow if paths are actually predictable.
Correct answer: D
Explanation: Agent loops need bounded autonomy and clear termination criteria. If paths are predictable, a workflow is simpler.
Why not the others: A allows more spend. B may not stop looping and costs more. C breaks tool usage.
Architectural principle: Bound agent autonomy with budgets and stopping conditions.
Ref: Building effective agents
Q13 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Question (Select TWO): Which are valid reasons to prefer a predefined workflow over an agent? A. The steps are well understood in advance and rarely vary. B. The team wants the model to choose its own tools freely. C. Predictable cost, latency, and easier testing matter more than flexibility. D. The problem is open-ended and the number of steps is unknown. E. The team wants to avoid writing any orchestration code. Correct answers: A, C Explanation: Workflows shine for well-defined tasks needing predictability and testability. Why not the others: B and D describe agent-suited conditions. E is false — workflows are orchestrated in code. Architectural principle: Choose workflows for predictability; agents for open-ended flexibility. Ref: Building effective agents
Q14 · D1 · Objective: Design multi-agent systems and orchestration strategies · Difficulty: Advanced
Scenario: A single research agent's context fills with raw search results and answer quality degrades after many steps. Question: Which redesign most directly addresses this? A. Increase temperature to diversify results. B. Delegate searching to worker agents with their own contexts that return concise, structured findings with sources to an orchestrator. C. Remove citations to save tokens. D. Ask the agent to "remember everything" via the system prompt. Correct answer: B Explanation: Context isolation keeps the orchestrator focused while workers handle noisy exploration. Why not the others: A doesn't address context size. C hurts verifiability. D can't expand capacity. Architectural principle: Use subagents for context isolation and structured handoffs. Ref: Building effective agents; Prep Lesson 1
Q15 · D1 · Objective: Design multi-agent systems and orchestration strategies · Difficulty: Advanced
Scenario: In a multi-agent design, all worker agents share one administrator credential for CRM access. Security objects. Question: What is the best response? A. Add a note in each worker's prompt to avoid destructive actions. B. Rotate the shared credential weekly. C. Give each worker its own identity and only the tools/scopes its task needs, enforced outside the model. D. Replace workers with a single agent using the same credential. Correct answer: C Explanation: Each worker is a trust boundary; least-privilege identities limit blast radius. Why not the others: A relies on prompting; B leaves over-broad access; D concentrates risk. Architectural principle: Isolation boundaries are security boundaries — least privilege per worker. Ref: Exam Guide D3/D5; MCP authorization spec
Q16 · D1 · Objective: Design multi-agent systems and orchestration strategies · Difficulty: Intermediate
Scenario: A coding assistant must implement a feature that touches an unknown number of files; the needed subtasks only become clear after reading the codebase. Question: Which pattern fits best? A. Orchestrator-workers: the orchestrator discovers subtasks dynamically and delegates to workers. B. A fixed three-step prompt chain. C. Static routing to one of three templates. D. A single prompt with the entire repository pasted in. Correct answer: A Explanation: Subtasks can't be predefined; orchestrator-workers handles dynamic decomposition. Why not the others: B and C assume known steps; D ignores size and focus limits. Architectural principle: Dynamic decomposition → orchestrator-workers. Ref: Building effective agents
Q17 · D1 · Objective: Design multi-agent systems and orchestration strategies · Difficulty: Expert
Scenario: Three parallel worker agents update a shared project plan and sometimes produce conflicting edits, leaving the plan inconsistent. Question: Which fix is most robust? A. Instruct workers to "be careful about conflicts." B. Run them sequentially without any other change. C. Let the last writer win. D. Make workers return typed proposals with explicit ownership of sections; a single reconcile/commit step applies changes and resolves conflicts. Correct answer: D Explanation: Clear ownership, structured handoffs, and a single writer remove race conditions while keeping parallelism for the work itself. Why not the others: A is not enforcement; B loses parallel benefits and doesn't define ownership; C silently drops information. Architectural principle: Parallelize work, serialize commits through a single authority. Ref: Building effective agents
Q18 · D1 · Objective: Apply decomposition techniques for complex problem solving · Difficulty: Intermediate
Scenario: Generating a 40-page RFP response in one call yields uneven quality and inconsistent terminology. Question: Which decomposition is best? A. Generate the whole document in one call with a longer prompt. B. Generate sections with shared style/terminology context and requirement mappings, then run an assembly and consistency pass with validation. C. Generate each section with unrelated prompts and no shared context. D. Reduce the RFP to a one-page summary. Correct answer: B Explanation: Section-level focus plus shared constraints and a consistency pass improves quality and coherence. Why not the others: A doesn't fix the issue; C causes inconsistency; D changes the deliverable. Architectural principle: Decompose along natural seams but retain shared context and a reconciliation step. Ref: Prep Lesson 1
Q19 · D1 · Objective: Apply decomposition techniques for complex problem solving · Difficulty: Advanced
Scenario: A pipeline was split into 12 tiny prompt steps. Latency tripled; steps don't use different tools or need validation between them; accuracy didn't improve. Question: What is the best change? A. Add more steps. B. Move every step to the largest model. C. Merge steps that share tools and context and have no intermediate validation need; keep splits where validation, tools, or specialization differ. D. Remove validation entirely. Correct answer: C Explanation: Over-decomposition adds latency and cost without benefit; split only where it improves reliability. Why not the others: A worsens latency; B raises cost; D reduces reliability. Architectural principle: Decompose to improve reliability or control, not for its own sake. Ref: Building effective agents
Q20 · D1 · Objective: Apply decomposition techniques for complex problem solving · Difficulty: Advanced
Scenario: A single prompt that extracts facts from medical-device complaints and then assesses regulatory reportability makes subtle errors in both parts. Question: Which redesign is best? A. Separate extraction (schema-bound, validated) from reportability reasoning applied to the validated facts, with different evals for each stage. B. Ask for chain-of-thought and trust the output. C. Merge both outputs into free text for readability. D. Use few-shot examples only. Correct answer: A Explanation: Two-stage decomposition lets you validate intermediate facts and evaluate each stage independently. Why not the others: B doesn't add validation; C removes structure; D doesn't separate failure modes. Architectural principle: Decompose to create validation points and component-level evals. Ref: Prep Lesson 1; D4 evals
Q21 · D1 · Objective: Design end-to-end architectures (input → processing → output → feedback loops) · Difficulty: Intermediate
Scenario: An assistant has been live for three months. Users can thumbs-down answers, but nothing links ratings to traces, and the eval set hasn't changed since launch. Question: What is the architectural gap? A. The output stage. B. The input stage. C. The model tier. D. The feedback loop: feedback isn't tied to traces or fed into evals and improvements. Correct answer: D Explanation: A feedback loop must connect user signals to traces, label failures, and grow the eval set. Why not the others: The issues described don't involve inputs, outputs, or model selection. Architectural principle: Close the loop: feedback → traces → evals → gated improvements. Ref: Exam Guide §6 D1/D4
Q22 · D1 · Objective: Design end-to-end architectures (input → processing → output → feedback loops) · Difficulty: Advanced
Scenario: A PoC works for five internal users with a hard-coded key, no retries, and no logging. A rollout to 5,000 users is planned. Question: Which set of additions best represents moving to production? A. A bigger model and a longer prompt. B. User-scoped authorization, retry/timeout/fallback handling, observability, and eval gates on production-like data with rollback. C. A new UI theme and onboarding video. D. A second PoC with different data. Correct answer: B Explanation: Production readiness is about the system around the model: identity, reliability, observability, and release control. Why not the others: A doesn't address production gaps; C is cosmetic; D repeats the PoC. Architectural principle: PoC proves possibility; production adds identity, reliability, observability, evaluation, and ownership. Ref: Prep Lessons 1–2
Q23 · D1 · Objective: Design end-to-end architectures (input → processing → output → feedback loops) · Difficulty: Intermediate
Scenario: An ADR says "Use a small model for triage," but gives no way to know when to change. Question: Which addition most improves the ADR? A. A longer description of the model's training. B. A list of team members. C. Measurable reversal triggers (e.g., switch if triage accuracy < threshold on the regression set or if escalation volume exceeds a limit). D. A statement that the decision is final. Correct answer: C Explanation: Defensible decisions state what evidence would change them. Why not the others: A, B, and D don't provide revisit criteria. Architectural principle: Every decision records alternatives and reversal triggers. Ref: Prep Lesson 1
Q24 · D1 · Objective: Design end-to-end architectures (input → processing → output → feedback loops) · Difficulty: Advanced
Scenario: Engineering favors an agent; operations favors a workflow; both argue strongly in a design review. Question: What is the best way to resolve this? A. Document both options against agreed criteria (accuracy, latency, cost, risk, operability) with evidence from the eval set, then record the decision with alternatives and reversal triggers. B. Pick the option from the more senior person. C. Choose the more advanced technology. D. Defer the decision indefinitely. Correct answer: A Explanation: Evidence against agreed criteria makes the decision defensible and auditable. Why not the others: B, C, and D don't use evidence or criteria. Architectural principle: Decide with criteria, evidence, and recorded alternatives. Ref: Prep Lesson 1; ADR practice
Q25 · D1 · Objective: Align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs) · Difficulty: Intermediate
Scenario: A contact center wants after-call summaries to cut agent after-call work. A baseline of 4.2 minutes exists. Question: Which success measurement is best? A. Number of summaries generated per day. B. Token count per summary. C. Agent satisfaction survey only. D. Change in after-call work time versus baseline, paired with a summary-quality gate on a reviewed sample. Correct answer: D Explanation: Tie the metric to the business outcome and guard quality so speed isn't gained by poor summaries. Why not the others: A and B are activity measures; C lacks an objective outcome. Architectural principle: Measure business outcomes against a baseline with quality guardrails. Ref: Prep Lesson 4
Q26 · D1 · Objective: Select appropriate architectural patterns (workflow, agentic, augmented LLM) · Difficulty: Intermediate
Question (Select TWO): Which statements are correct? A. The augmented LLM — an LLM enhanced with retrieval, tools, and memory — is the basic building block of agentic systems. B. In an agent, the control flow is always fixed in code in advance. C. Workflows are always more capable than agents. D. Workflows orchestrate LLMs and tools through predefined code paths. E. An augmented LLM requires multiple cooperating agents. Correct answers: A, D Explanation: These match Anthropic's definitions in "Building effective agents." Why not the others: B is the definition of a workflow, not an agent; C is false (they differ in predictability vs. flexibility); E is false. Architectural principle: Know the vocabulary: augmented LLM, workflow, agent. Ref: https://www.anthropic.com/research/building-effective-agents