1 The exam and the professional mindset · CCAR-P

What the exam tests: whether you make the decision a senior architect would defend in a design review, not whether you remember an API parameter.

How Professional differs from Foundations

Foundations asks "can you build it correctly?" Professional asks "should it be built this way, for this business, under these constraints, and how will it be run for the next two years?" Three shifts follow from that.

From components to systems. Questions span several domains at once: an integration choice that creates a governance problem, or a cost fix that degrades evaluation coverage. From correct to best. Most items have two or three plausible answers. The right one is the most proportionate control, the least-privilege option, or the fix aimed at the actual root cause.

From code to people. About 21% of the exam (Domains 6 and 7) is about discovery, communicating trade-offs, handoff, and enabling teams.

Six principles that answer most questions

Start from the business outcome. Pick the answer that ties the design to a measurable result (handle time, cost per case, SLA), not the most technically impressive one. Use the simplest pattern that works. A single well-prompted call beats a workflow; a workflow beats an autonomous agent; one agent beats a multi-agent system. Add autonomy only when the task needs open-ended decisions.

Least privilege, always. Remove a capability before you monitor it. Preventive controls beat detective ones.

Fix the root cause the symptoms point to. If answers went wrong after a data refresh, look at retrieval, not the model.

Measure before and after. Every change in a well-run system is gated by an evaluation set and observed in production.

Humans own high-impact decisions. Irreversible, regulated, or high-value actions get a human checkpoint, sized to the risk.

How items are built

Items are short stems with a scenario, a constraint, and a question such as "which change best…" or "what should the architect do first?" Multiple-response items tell you how many to select. Distractors usually fall into four families:

Distractor family What it looks like Why it is wrong

Bigger hammer Switch to a larger model, add more agents Ignores the root cause; raises cost Detect, don't pre‐

Wrong order Build before discovery, scale before evaluation Right action, wrong time

Worked example: taking three items apart

The fastest way to learn the exam's logic is to watch the method from Chapter 10 run on realshaped items. Each walkthrough below shows the reasoning a strong candidate uses, including the options they rule out and why.

Item 1: the "first step" trap

A logistics company's leadership has approved a budget for "an AI agent that optimizes our operations." The architect has been asked to start next week. What should the architect do first?

A. Build a prototype agent with access to the routing system to show early value
B. Choose the most capable Claude model so the agent can handle any task
C. Run discovery with operations leads to identify a specific, measurable problem and its baseline
D. Set up an MCP server for every operational system so the agent has full context Step 1, the last line: "What should the architect do first?" This is a sequencing question. The right answer is the earliest correct action, not the most impressive one.

Step 2, the constraint: the goal is vague ("optimizes our operations"). Vague goals point to discovery. Step 3, the domain: Stakeholder Communication & Lifecycle (Domain 6). The deciding principle is discovery before design. Step 4, eliminate: A is "wrong order" (build before discovery) and grants write access to a critical system. B is "bigger hammer" with no problem defined. D is "wrong order" and also capability bloat. Answer: C. Every later decision (pattern, model, tools) depends on knowing the problem, its value, and how success will be measured.

Item 2: the "best control" trap

A procurement agent reads supplier emails and can create purchase orders up to any value in the ERP system. Security is worried that a malicious email could trick the agent into creating fraudulent orders. Which change BEST reduces this risk?

A. Add a system prompt instruction to ignore instructions contained in emails
B. Log every purchase order the agent creates and review the log weekly
C. Require human approval for purchase orders and restrict the agent to approved suppliers
D. Move the agent to a more capable model that is harder to manipulate Step 1: "BEST reduces this risk." A control question: rank by strength.

Step 2: the risk is indirect prompt injection leading to a high-impact action. Step 3: Governance (Domain 5) and Integration (Domain 3). Principle: preventive beats detective; limit the blast radius. Step 4: A is a prompt-only defense and is unreliable on its own. B is detective and slow: the money is already gone by the weekly review. D is "bigger hammer"; no model is immune. Answer: C. It removes the agent's ability to cause harm alone and narrows its scope. Note that C would still be improved by also logging, but logging alone is never the best answer when a preventive control exists.

Item 3: a multiple-response item

A RAG assistant for a law firm answers questions about internal precedent memos. Partners complain that answers often cite memos that have been superseded. Latency and model version are unchanged. Which TWO actions address the root cause?

A. Add document status and effective-date metadata, and filter superseded memos at retrieval
B. Switch to the most capable model tier
C. Increase the number of retrieved chunks from 5 to 20
D. Re-ingest on a schedule tied to the memo system and alert when the index lags behind it
E. Tell the model in the system prompt to prefer recent memos

Step 1: "Which TWO" and "root cause." Judge each option on its own.

Step 2: model and latency unchanged; symptom is stale content. The cause is in ingestion and retrieval.

Step 3: Integration (Domain 3) with an Evaluation (Domain 4) diagnosis flavor. Step 4: A: true, it makes staleness filterable. B: false, the model is not the cause. C: false, more chunks would bring in even more superseded memos. D: true, it keeps the index aligned with

The exam and the professional mindset

the source. E: weak; the model cannot tell what is superseded without metadata, and prompt wording is not a data fix.

Answer: A and D.

What the three items teach

Lesson Applies when you see

Discovery before design Vague goals, new initiatives, "first" questions Prevent, then detect Risks that a capability change can remove

Fix where the symptom starts Change in data, config, or traffic with the model unchanged Judge multiple-response options one at a time "Select TWO" or "Select THREE" items

Common traps

Choosing the newest feature because it is new. The exam rewards fit, not novelty. Reading "first" as "best overall." When a stem asks what to do first, pick the diagnostic or discovery step.

Treating cost and quality as separate. Most good answers improve one without hurting the other.

Check yourself

A stakeholder wants a five-agent system for invoice extraction from a fixed template. What do you propose, and why?

Name two preventive controls and one detective control for an agent that can send email.

What does "criterion-referenced" scoring mean for how you should prepare?

Which domains have no counterpart on Foundations, and what share of the exam do they carry?

Answers: (1) A single structured-output call or a short deterministic workflow; the task is fixed and does not need autonomy. (2) Preventive: restrict recipients to an allowlist, require human approval to send. Detective: log and alert on outbound volume. (3) You must reach a fixed standard, not beat other candidates, so fix weak domains instead of polishing strong ones. (4) Governance (Domain 5), Stakeholder Communication & Lifecycle (Domain 6), and Developer Productivity (Domain 7); together they carry 35%.