CCAR-P - Solution Design & Architecture (17% of the exam) - Section 1.3

Select appropriate architectural patterns (workflow, agentic, augmented LLM).

Choosing between a single augmented LLM call (with retrieval, tools or memory), a predefined workflow where code controls the sequence, and an agent that directs its own steps. The discriminating judgment is that more autonomy costs predictability, latency and spend, so the simplest pattern that meets the requirement is usually correct.

augmented LLMworkflow with code-defined stepsagent with model-directed control flowautonomy versus predictability trade-off

Practice question for this objective

Free sampleSolution Design & Architecturehard

A health insurer's clinical coding assistant is a single model call augmented with retrieval over coding guidelines and a tool that looks up code descriptions. Its error rate on audited claims is 8 percent, and an engineer proposes rebuilding it as an agent that can plan, search repeatedly and check its own work. Review of the 240 miscoded claims finds that in 205 of them the retrieved guideline passages came from the previous year's codebook edition, which the index still holds alongside the current one, and that the model applied those passages faithfully. What does this evidence indicate?

AThe single call has no chance to check its own work, so an agent that searches again would catch the superseded passages before answering.
BThe model is applying outdated guidance absorbed in training, so fine-tuning on the current codebook edition is needed to override that knowledge.
CThe model's reasoning is too weak for coding rules, so moving the same call to a larger model tier would reduce the share of misapplied passages.
DThe errors originate in retrieval returning the superseded edition, so restricting retrieval to the current edition fixes them without changing the pattern. Correct

Before adding autonomy, trace errors to their layer; a retrieval defect in an augmented call is fixed in retrieval, not by changing the pattern. In 205 of 240 errors (about 85 percent) the model correctly applied the passages it was given, and those passages came from a superseded edition still held in the index. The defect therefore sits in the retrieval augmentation, not in the model or the pattern. Restricting retrieval to the current edition removes the cause, whereas an agent would search the same mixed index and add cost and unpredictability.

Why A is wrong: This is tempting because self-checking agents can catch some errors. It is wrong because a repeat search would query the same index holding both editions and could return the same superseded passages, so the agent adds cost and variability without removing the source of the errors.

Why B is wrong: This is tempting because outdated model knowledge is a real failure mode. It is wrong because the review shows the outdated guidance came from retrieved passages, not from the model's training, so fine-tuning would leave the index serving the wrong edition.

Why C is wrong: This is tempting because a larger tier is a common response to accuracy problems. It is wrong because the passages were applied faithfully, so the reasoning is sound and the input is wrong; a larger model given the same superseded passages would produce the same codes.

Why D is correct: Correct. Most errors trace to the retrieval augmentation serving outdated passages that the model then applied correctly, so filtering the index to the current edition addresses the cause while the single augmented call stays as simple and predictable as before.

See more CCAR-P practice questions, answers explained.

Exam traps in Solution Design & Architecture

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Keep the agent and add a system prompt rule that it must call the four steps in order, with a worked example run included for reference
Why it is wrong: This is tempting because it is a small change and often nudges the agent towards the intended order. It is wrong because a prompt instruction is guidance the model may still depart from, so the path remains model-directed and validators still cannot guarantee two runs on one application match; it also leaves the variable call count and latency in place.
The model tier is too small to follow a four-step instruction reliably, so a larger tier would remove both the skipped steps and the score spread.
Why it is wrong: This is tempting because more capable models follow instructions more consistently. It is wrong because a larger model would still be choosing the path on each run, so the variation would shrink at most, whereas the policy requires a guaranteed order that only code-defined control provides.
Move the agent to the most capable model tier so that it plans each answer in fewer model calls
Why it is wrong: Fewer calls per conversation sounds like a cost saving, which makes this tempting. It is wrong because a more capable tier costs more per call, so any reduction in call count is unlikely to offset it, and the agent would still be planning steps that are already known for most traffic.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.