Claude Certified Architect — Professional

1-Day Before Exam Review Notes

Use this as a final refresher based on the concepts and decision patterns covered by the AI Pathway CCAR-P question bank. Revisit weak areas from your mock exam rather than trying to learn everything again.

Start CCAR-P Mock Exam

1. Solution Design & Architecture

Use fixed workflows when branches and stop conditions are known and repeatability matters; use agents when the path must emerge from intermediate findings.
Choose multi-agent orchestration when work can be decomposed into genuinely distinct specialties, toolsets, contexts, or useful parallel tasks. Coordination cost must be justified.
Make routing, exception gates, trust boundaries, and recovery paths explicit. Do not hide a multi-stage controlled process inside one monolithic model call.
Design for auditability: the same inputs and policies should lead reviewers to the same controlled branch where determinism is required.

2. Tool, MCP & Structured Interface Design

Give tools narrow, non-overlapping contracts and least-privilege capabilities; do not rely on prompt wording to prevent prohibited actions.
Use structured outputs/JSON Schema for syntactic guarantees, then validate business rules separately. Schema conformance is not semantic correctness.
Treat tool failures as typed operational states where possible: validation, permission, business-policy, and transient failures need different recovery behavior.
For MCP and external integrations, design authorization, credential scope, approval boundaries, and failure handling before production.

3. Agent Orchestration & State

Parallel workers are valuable when tasks are sufficiently independent and can be aggregated under one rubric; sequential work is better when later steps depend heavily on earlier findings.
Preserve stable state, identifiers, provenance, and idempotency across retries and handoffs. A retry must not accidentally repeat an irreversible side effect.
Use deterministic hooks or enforcement layers for non-negotiable controls; prompts provide guidance but not guaranteed enforcement.
Pass the context a worker actually needs. Do not assume isolated agents automatically share conversation state.

4. Evaluation, Trade-offs & Optimization

Evaluate on representative workload cases using agreed quality, latency, and cost measures rather than choosing by model size or intuition.
Eliminate dominated configurations that provide no measured advantage, then choose among efficient candidates using explicit business thresholds.
Maintain regression evaluations for important failure modes and rerun them when prompts, models, tools, or workflows change.
Separate automated metrics from human judgment where the task includes subjective quality or high-consequence decisions.

5. Reliability, Security & Recovery

Design least privilege, fail-closed behavior, bounded retries, timeouts, and recovery paths before deployment.
Preserve successful partial work when one dependency fails, but mark the overall outcome degraded when required evidence is missing.
Do not confuse operational failure with a valid empty result. Downstream decisions must know whether evidence was absent or unavailable.
For unattended automation, permissions and credentials must technically block out-of-scope actions; retrospective logs alone are not sufficient.

6. Stakeholder Communication & Lifecycle Management

Production handoff should identify owner and version, passed evaluations, monitoring expectations, and a known-good rollback version.
After launch, continuously triage feedback, reproduce material failures, add evaluation or monitoring coverage, and assign accountable owners.
Present decision records that connect evidence to trade-offs, assumptions, chosen thresholds, and revisit triggers.
Prioritize feedback by consequence and evidence, not simply arrival order or frequency; rare authorization or security failures can outrank common wording preferences.

7. Developer Productivity & Operational Enablement

Use Claude Code and automation with narrowly scoped capabilities appropriate to the repository and environment.
Enforce protected-repository controls deterministically with hooks/permissions where the rule must hold regardless of model behavior.
In CI remediation, allow only the read/edit/test capabilities needed; deny push, deploy, and unrelated secrets unless the workflow explicitly requires and authorizes them.
Operational tooling should improve reproducibility and diagnosis without weakening security boundaries.

High-value distinctions

Workflow vs agent: known sequence and gates favor workflows; uncertain paths driven by intermediate evidence favor agents.
Prompt instruction vs enforcement: use permissions/hooks for controls that must never be bypassed.
Structured output vs correct output: JSON Schema guarantees structure, not truth or business validity.
Retry vs idempotency: retries are safe only when repeated operations cannot duplicate irreversible effects.
Monitoring vs evaluation: pre-release evals test expected behavior; production monitoring and feedback expose real-world failures.
Rollback vs latest version: production should have a known-good recoverable version, not depend on unpinned automatic change.

Final 15-minute checklist

Review the domain headings and your weakest mock-exam areas.
Explain the distinctions on this page without looking at the notes.
Revisit mistakes from your most recent attempt; do not start a new topic.
Confirm current exam logistics and policies using the official certification source.
Stop early enough to rest; prioritize recall and judgment.

About these notes: These are independent AI Pathway review notes derived from the approved CCAR-P preparation question bank. They are not an official Anthropic study guide or passing standard.