Practice — CCAR-P Domain 1: Solution Design & Architecture (17%)

16 scenario-based MCQs. Answer key + explanations at the bottom.

---

Q1

You are designing a claims-intake system for an insurer. Every claim goes through the same three steps — extract fields from the submitted PDF, validate them against policy rules, and write a triage record — and each step has a deterministic pass/fail check. The business sponsor has asked for "an AI agent." Which architecture best serves the actual requirement?

A. An autonomous agent given document, validation, and database tools, free to sequence its own steps B. Orchestrator–workers, with an orchestrator deciding per claim how to decompose the intake work C. A prompt-chaining workflow (extract → validate → write) with programmatic gates between the steps D. Evaluator–optimizer, with a second model critiquing each extraction until it accepts the result

Q2

According to Anthropic's architectural definitions, what actually separates an "agent" from a "workflow"?

A. In a workflow, LLMs and tools are orchestrated through predefined code paths; in an agent, the LLM dynamically directs its own process and tool usage B. In a workflow, the LLM never calls tools; in an agent, the system becomes agentic the moment the LLM invokes at least one external tool C. In a workflow, a single model handles everything; in an agent, two or more models must coordinate and exchange intermediate results D. In a workflow, requests complete synchronously; in an agent, execution always happens asynchronously as a long-running background job

Q3

A team built a monolithic "operations copilot" whose single system prompt now contains conditional rules for six distinct business functions (billing, HR, legal, IT, procurement, facilities). Accuracy on the rarer functions keeps degrading each time a new rule is added. As the architect, what is the best structural fix?

A. Upgrade the handler to the largest available model so the growing rule set can be weighted reliably B. Split the system: a cheap classifier routes each request to a specialized handler with its own prompt and tools C. Convert every conditional rule into a few-shot example so all six rule sets survive in one prompt D. Rebuild the copilot as an autonomous agent so the model can decide for itself which rules apply

Q4

Your company must summarize and score 120,000 archived vendor contracts once, for a migration project. No user is waiting on results, the budget is tight, and every contract is scored against the same fixed rubric. Which end-to-end design is best aligned?

A. A single long-running agent loop that iterates through the contracts one row at a time B. Synchronous streaming Sonnet calls fanned out across fifty client-side worker threads C. Orchestrator–workers, with an Opus orchestrator assigning contracts to subagents dynamically D. The Message Batches API on Haiku, with the fixed scoring rubric in a cached system prompt

Q5

In a multi-agent research system, two worker agents write intermediate findings into a shared scratch document that both also read from, and the final report is assembled from that document. Reports have started containing contradictory claims. What design rule was violated?

A. Workers must always execute strictly one after another and never run concurrently B. Workers must not share mutable state; results flow up to the orchestrator, which integrates them C. Workers must run on the same model so that their outputs stay stylistically consistent D. Workers must each receive the orchestrator's complete message history so outputs stay aligned

Q6

A VP of finance asks for "AI across finance operations" and offers four candidate projects. As the architect translating this into a first Claude deployment, which candidate is the strongest choice?

A. An autonomous agent that initiates vendor payments automatically once invoices appear approved B. An open-ended chat assistant answering any finance question employees ask, with no defined scope or success metric C. A generation system that drafts quarterly earnings guidance for direct external publication D. Invoice field extraction with schema validation, provenance pointers, and human review of low-confidence rows

Q7

An internal policy assistant answers thousands of employee questions per day against a stable 300-page handbook (~150K tokens). Answers must quote the handbook, and both latency and per-query cost matter. Which context architecture is best?

A. Cache the full handbook as a stable system-prompt prefix and use retrieval to point the model at the sections each query should focus on B. Send the entire handbook uncached with every request, since the model's context window can fit the whole document C. Use classic top-K retrieval alone, inserting three chunks per query and skipping caching infrastructure entirely D. Keep the handbook inside the running conversation history so that subsequent turns can reuse it without re-sending

Q8

A document-processing pipeline has clean input handling, a validated extraction step, and structured output — yet stakeholder complaints about quality keep arriving, and each prompt tweak seems to fix one case while breaking another. As the end-to-end designer, what component is missing?

A. A larger model for the extraction step, since persistent quality complaints indicate capacity limits B. A second extraction pass run in parallel, with a consensus vote across the two results C. An evaluation dataset and baseline, so every prompt change is measured against known cases before shipping D. Streaming output, so users see partial results sooner and perceive higher quality

Q9

A team implemented orchestrator–workers for support-ticket triage: an Opus orchestrator reads each ticket, plans, and delegates to one of five fixed handlers (billing, returns, tech, sales, escalation). Costs are high and latency is unpredictable. What should the architect recommend?

A. Replace the planning orchestrator with a cheap classifier that routes tickets directly to the five handlers B. Keep the orchestrator in place but downgrade it to Haiku so the planning step costs less per ticket C. Merge the five specialized handlers into one prompt so each ticket needs only a single model call D. Add prompt caching to the orchestrator's system prompt and keep the current design unchanged

Q10

An autonomous migration agent rewrites code in a legacy repository. Its transcript shows confident progress reports, but many "completed" files fail to compile — and the agent never notices. Which design principle for autonomous agents was violated?

A. It should have used extended thinking so its internal reasoning about each rewrite was more careful B. It should have been given a larger tool set so it could act on the repository more flexibly C. Its system prompt should have stated the quality bar explicitly at both the top and the end D. It should have obtained ground truth from the environment — running the build and tests — at each step

Q11

Users ask a single, self-contained question about their account, which requires one lookup in a billing system and a grounded answer. The team is debating implementing a full agent loop. What is the appropriate architecture?

A. An agent loop, since the presence of a tool call makes the system agentic by definition B. A single augmented LLM call — one model invocation with the lookup tool and retrieval attached — with no loop C. Orchestrator–workers, with the billing lookup delegated to a dedicated subagent for isolation D. Prompt chaining: classify the question, perform the lookup, answer, then verify, in four separate calls

Q12

A multi-agent document-analysis system runs Opus for the orchestrator and Opus for all thirty scoped worker subagents. Output quality is fine, but per-run cost is unsustainable. What is the first architectural change to evaluate?

A. Collapse the entire system into one long-running Opus conversation to avoid duplicated context B. Cut the number of workers in half regardless of how the task decomposition was designed C. Move the scoped workers to Haiku while keeping the stronger model for orchestration and planning D. Move the orchestrator to Haiku while the workers stay on Opus to preserve output quality

Q13

An architect is writing the decomposition spec for an orchestrator–workers system. According to good multi-agent design, what must every subtask brief define?

A. Its input contract, output contract, and allowed tools, packaged as a self-contained brief B. A pointer to the shared memory location where all the workers coordinate their state C. The orchestrator's full reasoning trace so the worker has complete upstream context D. The model ID, temperature, and token limits configured for every other worker in the fleet

Q14

A legal team wants Claude to draft compliance clauses that must pass a written 40-point checklist before release; drafts typically need several revision rounds, and the number of rounds can't be predicted. Throughput is secondary to correctness. Which pattern fits best?

A. Prompt chaining with one fixed review step appended after the drafting call B. Parallelization: generate five candidate drafts concurrently and select the best one C. Routing: classify each clause type and dispatch it to a specialized drafting prompt D. Evaluator–optimizer: a generator drafts, an evaluator applies the checklist, and the loop repeats until it passes

Q15

A product manager wants a "fully autonomous shopping agent" embedded in the checkout flow, which has a hard 3-second latency budget and a strict conversion SLA. How should the architect respond?

A. Accept the request, but run the agent loop on the fastest available model so it completes within the latency budget B. Counter-propose a predefined workflow such as routing, since agentic systems trade latency and cost for performance this flow doesn't need C. Accept the request, but enforce a capped iteration count so the agent's runtime is bounded and can never run away in production D. Counter-propose orchestrator–workers so the plan can be parallelized across concurrent subagents and finish faster

Q16

A Claude-based extraction system feeds a financial reporting process. Regulators must be able to trace any figure in the final report back to the exact source passage it came from, across every processing step. Which design decision satisfies this?

A. Log the final report outputs and rely on the model to explain a figure's origin when an auditor asks B. Persist each run's full prompts and raw model responses so auditors can replay the generation and compare C. Thread provenance through the pipeline: every extracted value carries a source pointer propagated into the output schema D. Add an attestation step where a second model reviews the report and confirms figures match the sources

---

Answers

Q1: C. The steps are known at design time and each has a deterministic check — the definition of a workflow, and prompt chaining with gates is the matching pattern. An autonomous agent (A) trades latency and cost for flexibility this task doesn't need; orchestrator–workers (B) is for dynamic decomposition, but every claim follows the same plan; evaluator–optimizer (D) adds an iterative quality loop where a deterministic gate already suffices. The sponsor's word "agent" describes a wish, not the architecture.

Q2: A. Anthropic's definitions: workflows orchestrate LLMs and tools through predefined code paths, while agents dynamically direct their own processes and tool usage. (B) is the classic misconception — tool use alone doesn't make a system agentic; workflows use tools too. (C) confuses the distinction with multi-agent design, and (D) confuses it with execution mode.

Q3: B. This is prompt dilution: many same-priority conditional rules in one prompt lower the weight of each, and the fix is architectural — route to specialized handlers with tight prompts and scoped tools. A bigger model (A) is a band-aid that doesn't remove the dilution; few-shot conversion (C) keeps everything in one overloaded prompt; an autonomous agent (D) adds dynamism without removing the crowded rule set.

Q4: D. One-off bulk work with no latency requirement is the canonical Batch API case: 50% off input and output, async within 24h, with Haiku cost-aligned for fixed-rubric scoring and the stable rubric cached across requests. Synchronous fan-out (B) is the tempting runner-up but forfeits the 50% batch discount for latency nobody needs. A single agent loop (A) is slow and monolithic; orchestrator–workers (C) pays planning overhead when rows are uniform and independent.

Q5: B. Shared mutable state between workers is an explicit anti-pattern — it creates races and contradictions; results must flow up to the orchestrator, which integrates them into one coherent output. Sequencing (A) hides the symptom but keeps the broken data flow. Same-model workers (C) addresses style, not contradiction; sharing full history (D) destroys the context-isolation benefit that makes subagents work.

Q6: D. The strongest first deployment has bounded scope, clear evaluation criteria, schema validation, provenance for audit, and a human review gate — everything needed to measure and defend business value. The payment agent (A) takes destructive actions that require confirmation gates and is the riskiest possible starting point. The open-ended assistant (B) has no success metric to evaluate; external earnings guidance (C) removes the human review point exactly where the stakes are highest.

Q7: A. For a stable corpus with a variable query stream, the recommended hybrid caches the long document as a stable prefix (cache reads at ~10% of input cost, kept warm by high traffic within the TTL) and uses retrieval to focus attention — supporting exact quotes from anywhere in the handbook. Classic RAG alone (C) is the runner-up but makes retrieval quality the bottleneck when a missed chunk means a wrong answer. Uncached full-context (B) pays full price every call and suffers lost-in-the-middle weighting; conversation history (D) balloons per-session cost and doesn't persist across users.

Q8: C. The missing piece of the end-to-end loop is feedback: an eval dataset and baseline, so every change is measured — a "better" change for one input class may hurt another, which is exactly the fix-one-break-one symptom described. A larger model (A) is an unmeasured guess; a consensus pass (B) doubles cost without telling you whether quality improved; streaming (D) changes perception, not correctness.

Q9: A. The handler set is fixed, so this is a classification problem, not dynamic decomposition — routing with a cheap classifier gives one classify plus one handler call, with low and predictable latency. Downgrading the orchestrator (B) is the runner-up but keeps an unnecessary planning step: the architecture is overkill regardless of model. Merging handlers (C) reintroduces dilution; caching (D) trims cost but leaves the variable-latency planning loop in place.

Q10: D. Autonomous agents require ground truth from the environment at each step to assess progress — without it they hallucinate progress, which is precisely the confident-but-broken transcript. Extended thinking (A) improves reasoning but cannot substitute for external verification; more tools (B) expand action without adding feedback; prompt emphasis (C) can't tell the agent whether the code actually compiles.

Q11: B. One predictable step and one tool call is the augmented-LLM case: a single model invocation with retrieval and tools attached. Re-invoking an agent loop for a single tool call (A) is a named anti-pattern — pure overhead, latency, and cost — and tool use alone doesn't require a loop. A subagent (C) adds orchestration for nothing; a four-call chain (D) is the plausible runner-up but multiplies latency and cost when the single call already covers the whole task.

Q12: C. Workers are scoped and narrow, so they benefit most from Haiku's speed and cost, while orchestration — planning and integrating — is where stronger reasoning pays; this is the standard fleet-of-Haikus economics. Inverting it (D) puts the weak model exactly where reasoning matters most. Collapsing to one conversation (A) trades a working decomposition for a monolith; halving workers arbitrarily (B) breaks the decomposition rather than repricing it.

Q13: A. Each subtask must be tagged with an input contract, an output contract, and its allowed tools, delivered as a self-contained brief — vague briefs make workers wander. Shared memory (B) is the shared-mutable-state anti-pattern; the full reasoning trace (C) is noise that defeats context isolation — workers need the relevant facts, not the parent's history; other workers' runtime settings (D) are irrelevant to a scoped task.

Q14: D. Evaluator–optimizer is built for exactly this: clear evaluation criteria (the 40-point checklist), quality prioritized over speed, and an unpredictable number of revision rounds. Chaining with one review step (A) is the runner-up but caps revision at a single fixed pass, which the scenario says is not enough. Parallel drafts (B) buy variety, not convergence to the checklist; routing (C) solves specialization, not iterative quality.

Q15: B. Agentic systems explicitly trade latency and cost for task performance, so an open-ended loop cannot honor a hard 3-second budget — the architect's job is to translate the request into a predefined workflow that can. A faster model (A) still leaves an unbounded number of steps. An iteration cap (C) is the tempting runner-up, but it is a reliability primitive against runaway loops, not a mechanism that fits a multi-step loop into 3 seconds; orchestrator–workers (D) keeps variable, plan-dependent latency.

Q16: C. Auditability is a design property, not an afterthought: provenance threading propagates source IDs through every step, so every value in the output schema carries its pointer and a regulator can trace any figure to source data. Post-hoc model explanations (A) are reconstructions, not evidence. Replaying prompts (B) is the runner-up but proves what the model saw, not which passage produced a specific figure; second-model attestation (D) adds an opinion, not a traceable link.