Claude Certified Architect – Professional Practice Test

How to Study for the Claude Certified Architect – Professional Exam

Master the pattern vocabulary from Anthropic's "Building Effective Agents" — the design questions map scenarios to patterns: prompt chaining for fixed sequential steps with checkable gates; routing when inputs fall into distinct categories that a classifier can separate; parallelization by sectioning when one call carries two competing concerns, and by voting when high confidence justifies multiple calls; orchestrator-workers when subtasks can't be predicted; evaluator-optimizer when drafts measurably improve under critique against clear criteria. The meta-rule outranks them all: start with the simplest augmented LLM call and add orchestration only when measured results demand it, and reserve full agents for unpredictable paths with environmental feedback, plus sandboxed testing and human checkpoints before irreversible actions.

For integration (the biggest domain), think least privilege first — remove unneeded tools rather than instructing the model not to use them, propagate user-level authorization instead of confused-deputy service accounts, and expect capability-bloat questions where 80 overlapping tools degrade selection. Know RAG failure signatures: confident-but-wrong answers after a document refresh point to stale indexes (fix with event-driven incremental indexing and recency metadata); identifier queries failing under pure semantic search call for hybrid BM25+embeddings; multi-tenant stores need server-side tenant filters. For compliance, memorize the trigger words: PHI → BAA and HIPAA-eligible configuration; EU personal data → deletion pipelines for the right to erasure; U.S. federal → FedRAMP-authorized environment. Evaluation questions reward multidimensional criteria (quality plus latency plus cost), code-graded checks for closed answers, rubric-driven LLM judges for subjective ones, and held-constant A/B tests. And for stakeholder scenarios, the professional answer is always quantified tradeoffs and runbooks, never "comply silently" or "escalate to the board."