Free CCAR-P Practice Questions (With Explanations)

The CCAR-P is a 63-question, 120-minute proctored exam. The passing score is 720 on a 1,000-point scaled scale, which means there's no published raw-score cutoff, and aiming for "just above 720" is not a safe strategy.

Multiple-choice and multiple-response items are both in scope. Multiple-response questions always state how many answers to select. The exam is closed book, no AI assistants allowed, and no breaks once you've started.

10 free CCAR-P practice questions#

These are real questions from our 306-question CCAR-P bank - the same pool the full practice exam draws from, not simplified samples. Answer before you expand the explanation.

Question 1#

A security architect is reviewing an egress allowlist for an agent that calls external APIs. The allowlist currently grants access to a partner domain that exposes several undocumented endpoints. Which approach best reflects the security implications of this allowlist? (Select one!)

• A. Treat the allowlist as a performance optimization and rely on a larger model to reduce misuse.
• B. Treat the allowlist as a detective control and add logging to compensate for the granted access.
• C. Treat the allowlist as a risk reducer and add more domains to lower the chance of a breach.
• D. Treat the allowlist as a capability grant and audit every reachable function on each approved domain.

Show answer and explanation

Correct answer: D. Treat the allowlist as a capability grant and audit every reachable function on each approved domain.

An egress allowlist is a capability grant, not a risk reducer. Every function reachable through an allowlisted domain becomes part of the agent's attack surface and must be audited, even after the domain is approved. Adding more domains or using a larger model does not reduce that surface. Logging is only a detective control that records misuse after it occurs; it does not shrink the granted capability.

Question 2#

You are designing a security architecture for Claude Code. How do tool calls interact with the security proxies in the system? (Select one!)

• A. Tool calls are not inspected at all and rely on the model's judgment
• B. Tool calls are only inspected after they enter the model's context
• C. Tool calls route through proxies that enforce policy and can inspect return values
• D. Tool calls bypass proxies and go directly to the external API

Show answer and explanation

Correct answer: C. Tool calls route through proxies that enforce policy and can inspect return values

In Claude Code and Claude Cowork, tool calls route through proxies that enforce network and file policy and can inspect return values before they enter the model's context. This is a key security feature. The other options either assume tool calls bypass proxies or that the model directly accesses external APIs, which is not the case.

Question 3#

You are designing a tool that lists projects from a paginated API. The API can return hundreds of projects, but the agent typically only needs to find one specific project. What is the best design for the list tool? (Select one!)

• A. Return all projects in a single response so the agent has complete information
• B. Return the first page, a total count, and a cursor for fetching additional pages
• C. Return only the first page and omit the total count to keep the response small
• D. Return a search tool instead that queries by name and returns matching projects

Show answer and explanation

Correct answer: B. Return the first page, a total count, and a cursor for fetching additional pages

The question specifically asks for the best design of the list tool when dealing with a paginated API that can return hundreds of items. Returning the first page, a total count, and a cursor allows the agent to understand that more results exist and fetch them only if needed, preventing context bloat and wasted tokens. Returning all projects wastes context and slows the response, while omitting the total count leaves the agent blind to whether more pages exist. Although a search tool is a useful pattern for finding specific items, it represents a different tool design rather than an answer to how the list tool itself should be structured.

Question 4#

A developer is configuring an agent with the Claude Agent SDK and wants to understand the difference between allowedTools and the tools parameter. Which statement about allowedTools is correct? (Select one!)

• A. It disables all tools not listed, making them completely unavailable to the agent.
• B. It pre-approves the listed tools to run without prompting but does not restrict the agent to only those tools.
• C. It only affects prompting and does not change which tools are available to the agent.
• D. It restricts the agent to only the tools listed, preventing any other tool from being called.

Show answer and explanation

Correct answer: B. It pre-approves the listed tools to run without prompting but does not restrict the agent to only those tools.

In the Claude Agent SDK, allowedTools only pre-approves listed tools to run without prompting; it does not restrict the agent to only those tools. Unlisted tools remain available and fall through to the permission mode and canUseTool callback. To restrict availability, you must use the tools parameter. The other options either claim allowedTools restricts availability or that it only affects prompting without changing availability, which is incorrect.

Question 5#

An architect is reviewing a custom MCP tool that is marked readOnlyHint: true. What is the actual behavior of this annotation? (Select one!)

• A. The tool requires explicit user approval for every call.
• B. The tool is denied access to the filesystem entirely.
• C. The tool is automatically sandboxed to prevent any write operations.
• D. The tool can still write to disk; the hint only affects parallel execution with other read-only tools.

Show answer and explanation

Correct answer: D. The tool can still write to disk; the hint only affects parallel execution with other read-only tools.

Tool annotations are metadata, not enforcement. A tool marked readOnlyHint: true can still write to disk; the hint only controls whether the tool can be called in parallel with other read-only tools. It does not enforce read-only behavior.

Question 6#

What is the isolation scope of Claude Cowork's virtual machine? (Select one!)

• A. The VM runs as a container on the host
• B. The entire host filesystem is mounted into the VM
• C. Credentials are stored inside the guest machine
• D. Only the selected workspace and .claude folder are mounted

Show answer and explanation

Correct answer: D. Only the selected workspace and .claude folder are mounted

Claude Cowork runs inside a full virtual machine using the platform's vendor hypervisor. Only the user's selected workspace and .claude folder are mounted; nothing else on the host is visible, and credentials stay in the host's keychain. Mounting the entire host filesystem, using a container, or storing credentials in the guest would violate the isolation model.

Question 7#

A developer is building a tool that retrieves a list of records from a paginated API. The API can return thousands of records, but the agent typically needs only a few specific ones. Which response design best balances context efficiency with the agent's ability to find what it needs? (Select one!)

• A. Return the first page, a total count, and a cursor so the agent can fetch more only if needed
• B. Return all records in a single response so the agent never misses a match
• C. Return only the first page with no indication that more records exist
• D. Return a random sample of records to keep the response small

Show answer and explanation

Correct answer: A. Return the first page, a total count, and a cursor so the agent can fetch more only if needed

For paginated APIs, a list tool should return the first page, a total count, and a cursor so the agent can fetch more pages only if needed, rather than auto-fetching hundreds of items. Returning all records wastes limited context and is an anti-pattern. Returning only the first page without a cursor or total count hides the existence of further data, and returning a random sample risks missing the target record entirely.

Question 8#

You are building a customer support agent that needs to look up order status and issue refunds. The agent will have access to a database of orders and a payment processing API. You are designing the tool set. Which approach best balances capability with security and reliability? (Select one!)

• A. Create a single tool called manage_order with an action parameter that can look up status, issue refunds, and update shipping addresses, so the agent has one entry point for all order operations.
• B. Create separate tools for lookup, refund, and address update, but only include the lookup tool in the agent's tool set, and handle refunds through a separate approval workflow.
• C. Create separate tools for lookup, refund, and address update, and give the agent all three so it can handle any customer request without needing additional permissions.
• D. Create a single tool called manage_order with an action parameter, and include it in the agent's tool set along with a logging tool to track all refund operations.

Show answer and explanation

Correct answer: B. Create separate tools for lookup, refund, and address update, but only include the lookup tool in the agent's tool set, and handle refunds through a separate approval workflow.

Least-privilege design means the agent only has the tools it needs for its actual task. A support agent that primarily looks up order status should not have direct access to refund or deletion tools, as that widens the security surface and creates risk of misuse. Combining operations into a single tool with an action parameter reduces tool count but does not reduce the capability, the agent still has refund authority. Adding logging is a detective control that does not reduce the privilege. The correct approach is to scope the agent's tool set to only what it needs, and route sensitive operations through a separate approval workflow.

Question 9#

You are designing a tool for a Claude-powered customer database system. The tool search_contacts(query) returns a list of matching contacts with their IDs and basic metadata. The system also has a get_contact_details(contact_id) tool that returns the full record. During testing, the model frequently calls get_contact_details for every contact in the search results, even when only a summary is needed. What is the most effective way to address this behavior? (Select one!)

• A. Add a warning to the get_contact_details description stating that it should only be called for a single contact.
• B. Modify search_contacts to return a summary field for each contact so the model can answer most questions without calling get_contact_details.
• C. Add a limit parameter to get_contact_details so the model can only fetch a few records at a time.
• D. Remove get_contact_details entirely and include all details in the search results.

Show answer and explanation

Correct answer: B. Modify search_contacts to return a summary field for each contact so the model can answer most questions without calling get_contact_details.

Including a summary field in search results gives the model the high-signal information it needs for most queries, reducing unnecessary detail fetches. A limit parameter does not address the root cause of over-fetching. A warning in the description is a weak control that the model may ignore. Removing get_contact_details and including all details in search results would bloat the context window for every search, which is worse.

Question 10#

You are a solution architect at a customer support company. Your team is building an agent that helps customers with order issues. The agent has access to a tool that searches for orders by customer ID. When no orders match the search, the tool currently returns an error with the message 'No orders found.' You are concerned that this will confuse the agent. What is the most appropriate way to handle this case? (Select one!)

• A. Return a successful result with an empty results array so the agent knows the search succeeded but found nothing.
• B. Return an error with a generic message so the agent can retry the search.
• C. Return a successful result with a null value so the agent knows the search returned no data.
• D. Return an error with the message 'No orders found' so the agent knows the search failed.

Show answer and explanation

Correct answer: A. Return a successful result with an empty results array so the agent knows the search succeeded but found nothing.

A tool should distinguish a successful empty result from an error. 'No matches found' should be a successful result with an empty results array, not an isError result. This tells the agent that the search executed correctly but found no matching orders, which is a valid outcome. Returning an error would incorrectly signal a failure and might cause the agent to retry unnecessarily. Returning a null value is ambiguous and not a standard pattern. Returning a generic error message is also incorrect because the search did not fail.

How to Use This Sample Test#

120 minutes for 63 questions works out to roughly 114 seconds per question on average. That's enough time for a careful read, but not enough to re-read a 200-word scenario three times.

For this sample set, give yourself 19 minutes (10 questions × 114 seconds, rounded up slightly for transitions). Set a timer before you start. If you're consistently burning 3+ minutes on a single item, that's a pacing signal to address before exam day.

Don't memorize mechanisms. Practice scenario extraction. The exam is built around described systems, the skill it tests is pulling the decision-relevant constraint out of a long paragraph rather than recalling a definition.

Self-scoring:

• 7–10 correct: Ready to book. Focus remaining prep on Integration and Governance, where the reasoning patterns diverge most from intuition.
• 5–6 correct: One more timed session. Review every wrong answer for the reasoning failure, not just the right answer.
• Below 5: Domain gaps or reasoning framework mismatches. Identify which domains you dropped points in and work through those specifically before retaking.

Evidence from timed practice suggests building a comfortable margin above 720 before booking. The scaled scoring means you have no precise visibility into how raw performance maps to the final number.

What These Questions Reveal About the Real Exam's Style#

Coming into this from a non-AI background, say from software architecture or consulting, there are two things that land differently than expected.

The first is that scenario length varies dramatically. Some items are two sentences. Others are long-form system descriptions with constraints embedded in the middle of the third paragraph. The two-sentence items tend to test precise recall of how a mechanism works. The long-form items test whether you can extract the decision-relevant facts and discard the noise. Both types are on the exam; neither is harder in the abstract.

The second is that the Governance and Safety domain suspends the reasoning pattern that works everywhere else. For most domains, you can rank answer choices by how well they solve the stated problem efficiently and safely. The Governance domain tests a different priority order: legal and regulatory questions require the assessment to be performed before processing begins, and the rights of the affected person are in scope before the business goal. If you apply the general efficiency-first heuristic to a fairness assessment question: wrong answer. The pattern across pass reports shows this catches candidates off guard even when they've studied the domain.

Integration is the hard domain on this exam. At 19%, it's the largest, and it's built around a specific kind of reasoning: what should happen at each boundary between components. The reasoning pivots on three decision points: retrieval matching (small matches are precise; big matches ground results), contract validation (must be deterministic at team boundaries), and data access (request a verdict from another team, don't import their tools). Each represents a hard design constraint, not a preference.

The Solution Design domain has a trap worth naming: adding a pipeline stage is almost never the credited answer. Complexity has to be justified by explicit constraints in the scenario. If the scenario doesn't give you a number or a stated requirement that forces an additional stage, the correct move is usually to remove one or consolidate.

Domain Breakdown#

Domain 1, Solution Design & Architecture (17%)#

17%

This domain tests whether you can justify complexity, not just design it. The key question in every scenario: does the architecture earn its stage count? Multiple experience threads say the same thing, adding a pipeline stage is almost never the right answer. The credited move is usually to remove or consolidate. Official documentation covers workflow versus agent design choices, system boundary definition, and architecture simplification.

Domain 2, Claude Models, Prompting & Context Engineering (13%)#

13%

The smallest of the core domains at 13%, but candidates who skip it pay for it in the Integration questions that build on it. Know what platform primitives already solve before reaching for a custom workaround. For prefix caching, dynamic values belong after static content; conversation history is a growing stable prefix. Don't overcomplicate context assembly when a simpler structure works.

Domain 3, Integration (19%)#

19%

Integration is the hard domain. Full stop. It's the largest at 19%, and some candidates who've written up their experience report underestimating it more than any other domain. The official blueprint covers retrieval design, tool and system boundaries, handoffs between components, machine-to-machine contract validation, and cross-team data routing. The reasoning framework: match small and ground big on retrieval, validate contracts deterministically at machine-to-machine boundaries, and when you need another team's judgment over data only they hold, request a verdict rather than importing their tools.

Domain 4, Evaluation, Testing & Optimization (16%)#

16%

Label quality is the recurring theme here. An eval is only as trustworthy as the agreement between the people who labeled it. What shows up repeatedly in community threads: statistical traps, particularly many-arm comparisons where at least one false winner is close to expected. The domain tests whether you know when a change actually improved performance versus when the numbers are noise.

Domain 5, Governance, Safety & Risk Management (14%)#

14%

This is the domain where the general answer-ranking heuristic breaks. Legal questions want the assessment done before processing begins. Rights of the affected person come before the business goal. Expect fairness assessment scenarios. Those who failed and retook it say this domain is the most common source of avoidable misses, not because the material is obscure, but because the priority order is counterintuitive.

Domain 6, Stakeholder Communication & Lifecycle Management (14%)#

14%

Conflicting stakeholder requirements have a specific credited resolution: make the conflict visible and facilitate a decision. The exam does not credit documenting both sides, picking one, or inventing a compromise. The domain also covers ROI defense and net unit economics, lifecycle stage identification, and discovery. At 14%, it ties with Governance for the third-largest domain.

Domain 7, Developer Productivity & Operational Enablement (7%)#

7%

The smallest domain at 7%. Official documentation lists this domain in the blueprint; limited community data exists on its specific question patterns. Don't spend disproportionate prep time here. If you're tight on time, drill Integration and Governance first.

What the Exam Reveals About Hands-On Readiness#

This exam has a lab-heavy reputation. Actually, that's not quite right. The exam tests judgment you only develop by building.

Everything is multiple-choice and multiple-response, so there's no terminal session. But the questions are written from the perspective of someone who has made real architecture decisions, not someone who has read about them.

Build a two-stage Claude pipeline for Solution Design, then simplify it to one stage. Document what you gave up and what you gained. That exercise surfaces the complexity justification reasoning the exam tests directly.

For Integration, wire a Claude-powered function to an external API with structured output validation, then wire a second Claude agent that calls the first one and needs a verdict, not raw data. The distinction between "I need their tools" and "I need their judgment" is clearer after you've built both. That distinction is what Integration questions probe.

For Evaluation: label a small set of outputs yourself, then label them again two weeks later and measure your own agreement. For Governance: before you build the happy path in any scenario, write out what the failure mode looks like and who bears the cost.

Cost and Value#

The CCAR-P costs $175. The cert signals demonstrated competency within the Claude Partner Network, which is a gating criterion for partner organizations to field consultants on Claude implementations.

The demand context is real. As of October 2026, more than 175,000 Claude certifications had been issued across 46,000 firms, with $100M invested in training 10,000 Frontier Deployed Engineers by end of 2027. Organizations delivering Claude implementations need certified practitioners; the CCAR-P sits at the top of the practitioner track. At $175 for the exam, the direct cost is low relative to professional certification norms. The meaningful investment is prep time, and since there's no official practice exam, that time requires deliberate sourcing of practice material.

Frequently Asked Questions#

Is the CCAR-P hard? It depends on your background. The exam doesn't test Claude knowledge in isolation; it tests system design judgment applied to Claude-based architectures. Candidates with strong software architecture experience tend to find Solution Design and Integration manageable, but Governance trips people up regardless of background because the priority order is counterintuitive. Community experience suggests Integration is the domain most candidates underestimate.

Do I need to be at a partner organization to take the CCAR-P? Yes. Registration requires a partner company email address. The exam is only available to employees of Claude Partner Network organizations. Personal email addresses will not work. If your organization isn't in the partner network, there's an application process at the Claude Partner Network page before you can register for the exam.

What happens after I pass? A score report appears on screen immediately. A digital Credly badge is issued on a pass. The credential is valid for 12 months from the award date. Renewal is available through an unproctored assessment if done within the validity window. The score report includes your scaled score and percent-correct by domain, which is useful for identifying where to focus if you retake.

Is there an official practice exam? No. Anthropic does not publish an official practice exam for the CCAR-P. The official exam guide lives behind the partner portal. CertCompanion's CCAR-P practice bank is one of the few places to find questions mapped to the current blueprint with full explanations. Start with 30 free questions at /exams/anthropic-claude-certified-architect-professional-ccar-p.

What's the retake policy? No mandatory waiting period between attempts has been published by Anthropic. Book a retake through Pearson VUE the same way as the initial registration.

What's the difference between CCAR-P and CCAR-F? The CCAR-F (Claude Certified Architect Foundations) is the entry-level architect certification at $125. The CCAR-P (Professional) is the advanced tier at $175, with a more complex domain structure and scenario-based questions that require deeper judgment. Both count toward Claude Partner Network eligibility for partner organizations.

What's the best way to prepare without partner portal access? Focus on the domain breakdown and task statements available through public community resources. Build the hands-on exercises described in this post. Use CertCompanion's practice questions to calibrate your reasoning against exam-style scenarios. The free Claude Academy courses cover foundational concepts and are open to anyone without a partner requirement. Browse every Anthropic certification practice exam to compare the tracks.

Should I take CCAR-F before CCAR-P? There's no official prerequisite. Some candidates sit CCAR-P directly. Those who've written up their experience report the Foundations exam is useful for calibrating familiarity with how Anthropic frames architecture decisions, but it isn't required. If you have a strong software architecture background, CCAR-P directly is a reasonable path.

After You Pass#

The CCAR-P is the top credential in the Claude practitioner track. Roles that list it or its equivalent include Frontier Deployed Engineer, Claude Certified Architect, and consultant or partner practitioner positions at Claude Partner Network organizations.

The credential's value is tied to the Claude Partner Network ecosystem: partner organizations need certified architects to staff Claude implementations, and the CCAR-P is the benchmark credential for that role.

For next steps, the Claude Certified Developer Foundations (CCDV-F, $125) is worth holding if you're building toward a developer-focused track alongside architecture work. The Claude Frontier Deployed Engineer designation is a separate, nomination-only program requiring organizational sponsorship; it's not an exam path, but it's the next tier of recognition for architects doing production Claude work.

Renewal is free if done within the 12-month validity window via an unproctored assessment. Don't let it lapse; the renewal path is the lowest-friction option.

The CCAR-P doesn't reward knowing what Claude does; it rewards knowing how to design systems that use Claude well. The difference between those two things shows up in every practice question you get wrong before exam day. Use this sample set to find the gaps now: drill the domains where you dropped points, retake the sample until the reasoning clicks, and schedule the exam when you're consistently hitting 8 or above. Start with 30 free practice questions and full explanations at CertCompanion's CCAR-P exam page.

Checked against official Anthropic CCAR-P exam documentation and current candidate reports. Last verified 2026-10-15.