Claude Certified Architect - Professional Practice Test

Question 1 (Integration). A bank is deploying a Claude-based agent that retrieves customer account balances from the core banking API during support conversations. The security team requires that API credentials never appear in prompts, logs, or conversation transcripts. What is the correct integration pattern?

• A. Include the API key in the system prompt, marked "do not repeat this"
• B. Route external calls through MCP server tools with scoped, server-side credentials
• C. Paste the customer's account data into the user prompt so the model can reason over it
• D. Let the model retrieve balances from its training knowledge

Correct answer: B. MCP tools execute server-side with scoped credentials the model never sees — the model invokes the tool by name and receives only the result, so secrets stay out of the context window entirely. Option A fails the requirement: system prompt contents routinely end up in logs and can leak through prompt-injection or output quoting. C is a data-exposure problem, not a solution. D is a hallucination trap — the model's training data knows nothing about live account balances. The exam tests this pattern repeatedly: secrets belong on the server side of tool calls, never in context.

Question 2 (Evaluation, Testing & Optimization). You are about to launch an agent that summarizes insurance claim denials for customer support staff. Before launch, how do you establish that its outputs are good enough?

• A. Launch to a small pilot group and monitor satisfaction scores
• B. Build a golden dataset of claim-denial cases with expert-rated expected summaries, measure pass rates plus human spot-checks, and run a red-team set probing for PII leakage and fabricated policy citations
• C. Ask the model to rate the quality of its own summaries
• D. Check that the model's confidence scores are high across a validation set

Correct answer: B. Professional-grade evaluation means defined ground truth (the golden dataset), a repeatable pass-rate metric, human spot-checks for judgment calls, and a separate adversarial set for the failure modes that matter most — here, leaking PII and inventing policy language. A alone skips pre-launch evidence. C is circular — a model grading its own homework. D confuses token-level confidence with output correctness. Expect the exam to punish any evaluation plan that lacks a golden dataset and adversarial coverage.

Question 3 (Governance, Safety & Risk). A wealth-management firm is deploying a Claude agent that drafts client-facing investment summaries. Regulators require that every client-facing output is reviewable and that the firm can prove its controls were active. Which design satisfies this?

• A. A detailed system prompt instructing the model to stay within policy
• B. Layered controls: system prompt constraints, output classifiers/guardrails, human-in-the-loop approval before any client-facing output, and full audit logging of inputs, outputs, and control decisions
• C. Fine-tuning the model on the firm's compliance handbook
• D. Restricting the agent to internal staff only

Correct answer: B. Regulated deployments require defense in depth: the prompt sets behavior, classifiers catch what the prompt misses, a human approves what reaches the client, and audit logs prove all of it after the fact. A alone is a single point of failure — prompts don't self-verify. C changes model tendencies but provides no runtime enforcement or evidence. D avoids the problem instead of solving it. The exam's governance questions consistently reward layered, auditable controls over any single mechanism.

Question 4 (Solution Design & Architecture). An underwriting team wants an agent that reviews loan applications: it must pull applicant data from three internal systems, apply policy rules, flag edge cases for human review, and produce a written rationale. Which architecture is most appropriate?

• A. One comprehensive prompt that instructs the model to complete all steps in a single call
• B. Decompose the workflow into orchestrated steps with defined handoffs — data gathering, rule application, edge-case triage, rationale generation — using structured outputs between steps and explicit checkpoints
• C. Fine-tune a model on historical underwriting decisions and let it handle applications end to end
• D. Build a deterministic rules engine and skip the model entirely

Correct answer: B. Multi-step, high-stakes workflows are where agentic decomposition earns its keep: each step is inspectable, outputs are structured and validated between handoffs, and the human-review checkpoint is a natural seam in the design. A single mega-prompt (A) is untestable — one failure corrupts everything downstream with no visibility. C trades away explainability, which underwriting requires. D ignores the unstructured-data reality (free-text applications) that makes the model useful. The professional exam expects you to reach for orchestration with checkpoints whenever the workflow is long, consequential, and auditable.

Question 5 (Claude Models, Prompting & Context). Your agent works against a large codebase (~150k tokens). Each conversation turn currently resends the full context, driving up latency and cost. Which approach best reduces context overhead without losing needed information?

• A. Send the full context every turn for maximum accuracy
• B. Use prompt caching for the stable repository context, retrieve only the relevant files per task (RAG), and summarize intermediate reasoning between turns
• C. Fine-tune a model on the codebase to eliminate context needs
• D. Switch to a smaller model with a smaller context window

Correct answer: B. Stable content (the repo) is exactly what prompt caching is for — it slashes cost and latency on repeated prefixes. RAG keeps each turn's working set to the relevant files instead of the whole codebase. Summarizing intermediate state prevents context bloat across long sessions. A wastes money and tokens on every turn. C doesn't remove the need for working context and is wildly disproportionate for this problem. D reduces the window instead of managing it. Context economics — caching, retrieval, summarization — is a recurring professional-tier theme.