Anthropic CCA-F: Claude Certified Architect Foundations Exam Guide

Claude Certified Architect – Foundations is an architecture exam, not a test of how many prompt tricks someone can remember. Its questions ask whether you can design an application that uses Claude with appropriate boundaries: which work should be deterministic, when an agent earns autonomy, how a tool call should be authorized, what belongs in a working context, and which evidence makes a production decision defensible. That distinction should shape preparation from the first study session.

The qualification appears on this site’s commercial listings under the shorthand CCA-F. Anthropic calls the credential Claude Certified Architect – Foundations; reproduced versions of its July 2026 guide identify the exam code as CCAR-F. Use the identifier and policies displayed in Anthropic’s current registration material as authoritative. The Claude Certified Architect – Foundations exam listing is a separate study-product destination; it does not issue Anthropic’s certification or determine who can register.

What Anthropic actually tests

Architect Foundations evaluates the decisions behind a Claude-powered system. The candidate must be able to explain why a support assistant should call a narrow order-lookup function instead of querying unrestricted customer data, why a model-generated tool request still needs application-side permission checks, and why a valid structured response may still contain the wrong conclusion. In each case the architecture is evaluated as a whole, not by looking at a model prompt in isolation.

Anthropic’s current certification description places the credential at the Foundations level of the Architect role. An architect may decide deployment choices, agentic versus single-shot design, evaluation, costs and responsible delivery. The differences among the four Claude certification tracks come down to responsibility: Developer Foundations emphasizes implementing Claude applications, Associate Foundations emphasizes effective everyday use, and Architect Professional covers wider enterprise design and governance beyond the Foundations architect’s scope. The Architect Professional credential comparison is useful when a candidate is deciding which architectural responsibilities match their current work, not as a claim that one exam is a prerequisite for the other.

Exam format, price, eligibility and score

Anthropic Partner Academy currently lists 60 questions in 120 minutes, in English, with multiple-choice and multiple-response items. Delivery is online proctored or through a Pearson test center. The listed examination fee is US$125 before any applicable partner discounts, and the credential is valid for 12 months from the award date. These are issuer examination terms, not the pricing or validity of any third-party practice material. Check the Academy again when booking because the testing program can change.

The minimum reported passing score is 720 on a scaled range of 100–1,000. It does not follow that a candidate needs 72% of the questions correct, nor can the score be converted reliably into an allowed number of wrong answers. Anthropic explains that scaled scores permit fairer comparison across forms of different difficulty. A score report also provides section-level performance that can direct later study.

There is an important access restriction. Anthropic Partner Academy says Claude Certification is currently available through the Claude Partner Network and requires an eligible partner work email; personal emails and non-partner registration are not accepted. Candidates must be at least 18. This is separate from any recommended technical experience or existing certifications. There is no formal Architect Foundations exam prerequisite listed in the Academy’s program FAQ. The candidate who can study this subject is not necessarily eligible to book the official examination, so verify partner status before paying for an exam-specific preparation plan.

If a candidate fails, Anthropic’s published program FAQ specifies increasing waits before retakes—14 days, then 30, then 90—and a maximum of four attempts per rolling twelve months for an individual exam. Retakes are charged at the applicable exam rate, subject to the issuer’s partner discount policy. A candidate who has passed should track their certification expiration separately from any practice subscription. The Academy currently describes a non-proctored renewal assessment for eligible on-time renewal and a full examination after lapse; always follow the latest email or portal instructions for the specific credential.

The five weighted domains

The examination blueprint is not an equal-sized reading list. Agentic architecture carries the largest share, but two separate 20% domains test Claude Code operations and prompt/output design. The table below preserves Anthropic’s published domain names and weighting. The third column describes what a useful architecture exercise might involve; it is editorial teaching guidance rather than a quotation of hidden examination questions.

The weighting is a sensible allocation guide, not permission to abandon a smaller domain. If a candidate cannot distinguish a tool result from a tool instruction, they can fail reasoning in several domains at once. Cross-domain concepts such as least privilege, idempotency, test evidence and human authorization should be learned through complete scenarios rather than memorized separately as slogans.

Design the study work around one realistic system

Consider an internal support service that gathers product-policy evidence, reads an order history, drafts a response and proposes an adjustment. Begin by writing down what the service is allowed to accomplish. It may fetch records under the current representative’s identity, and it may draft a refund proposal; the refund itself requires human confirmation. Now mark which steps are deterministic and which require flexible reasoning. Identity verification, monetary thresholds and final authorization belong in application code. Choosing between several relevant policy passages may be model-assisted, subject to verification.

The same application can become your running example for all five domains. In agentic architecture, decide whether it needs a loop and where it stops. In tool design, specify a minimal read-only account lookup and structured failure reports. In Claude Code, decide what a development assistant can change in the project and what requires review. In structured output, define a response with source references and an explicit cannot-decide state. In context management, determine what should be retrieved for each case and how expired policy documents are excluded. A candidate who can defend those decisions learns more than someone who has highlighted every page of the API reference.

For scenario study, begin with the user outcome, permitted actions, trusted facts, cost or latency requirements, irreversible operations and the evidence needed to justify a decision. In Architect Foundations scenarios involving agents, MCP and context, compare possible approaches against those constraints rather than choosing a technology because it sounds more capable. A narrowly bounded workflow can be the stronger design when actions must be repeatable, auditable and easy to recover.

A four-week preparation plan you can adapt

A four-week schedule is practical for an experienced engineer or architect who can set aside focused sessions; it is not an official Anthropic study duration. A candidate who has never used tool calling or Claude Code should allow more time for implementation. Rather than splitting study hours precisely by percentages, use the domain weights as a guide and shift extra time toward demonstrated weaknesses.

Week one: establish the architecture baseline. Read the current examination outline, then build the smallest bounded workflow that solves your chosen case. First make a deterministic sequence, then introduce an agent only at the point where the path genuinely depends on new evidence. Trace `tool_use` and `tool_result` exchanges, record normal completion and interruption states, and decide which side effect must not be retried without checking the system of record. Draw the coordinator/subagent boundary and specify the context passed across it. Finish the week by explaining why one solution has fewer failure states than its more autonomous alternative.

Week two: defend every tool boundary. Implement or at least specify two narrowly scoped tools: one that reads an authorized record and one that would propose a business action without committing it. Define inputs, returned provenance, typed error responses and an application-side authorization rule. Rehearse timeouts where an action might have happened, a tool result containing an embedded malicious instruction, and an MCP server whose discovery metadata is convenient but not independently trustworthy. Your test is whether you can describe a safe recovery path without asking the model to invent permissions.

Week three: make the development and output path dependable. Study how `CLAUDE.md` communicates project guidance, when scoped rules or skills load, what plan mode should inspect before execution, and which operations need actual permission settings or hook enforcement. The Claude Code rules and CI analysis illustrates why guidance and enforcement are not the same. In parallel, build a prompt that extracts a support decision into a documented schema. Confirm what the schema validator can prove; then deliberately supply conflicting records and confirm that your system can refuse to invent a conclusion.

Week four: practice under uncertainty. Revisit the five-domain table using short scenarios you write yourself. For each answer, record the first decisive constraint and the reason the nearest alternative fails. Test long sessions, stale retrieved passages, conflicting sources, partial tool failures, and escalation thresholds. Simulate time pressure on a set of original problems; then use the errors to choose the next technical exercise. Do not memorize sample answer letters or treat a practice percentage as an official scaled passing score.

Five exercises that expose weak understanding

The ambiguous refund: A support case mentions two accounts, but the user is authorized for only one. What is the safe next tool call? The right design identifies the account and verifies permission before retrieval. A confident model guess is not identity resolution. Document what the assistant should say when it cannot resolve the ambiguity.

The timed-out mutation: A tool that can create a case note times out after submission. Retry immediately, or query whether the note exists? A production design uses an idempotency key or a reconciliation read when the operation may have committed. The architecture must distinguish transport failure from business failure and must not turn uncertainty into duplicate side effects.

The untrusted document: A retrieved customer policy includes text commanding the assistant to override approval rules. Treat that text as evidence to summarize, not higher-priority instructions. The policy document cannot confer action privileges, and logs should record the attempted instruction without blindly executing it.

The valid but wrong JSON: A response matches your schema and includes a price, source ID and approval flag; the price came from an expired tariff. Structure is correct, business validity is not. Test source freshness and approval rules independently, and require a recoverable missing-evidence state.

The bloated session: An agent carries dozens of stale cases and tool traces forward. More context can reduce focus and leak unnecessary information. Establish a bounded working set, fetch only the case-relevant records, compact when appropriate, and retain enough references to reconstruct why a decision was made.

How to tell whether you are ready

Readiness is not simply how many PDFs you have read. A useful candidate can draw an agent loop, specify a tool contract, explain what a hook can enforce, distinguish schema success from factual accuracy, and identify when the architecture must stop and ask a person. They can also compare two architectures without assuming the most elaborate design is best.

Check five kinds of evidence before booking: you have built or closely reviewed one relevant Claude workflow; you can explain its permissions and failure recovery; you can defend where context comes from; you can turn output quality into measurable acceptance criteria; and you can solve a scenario without relying on undocumented exam leaks. A weak result in one domain is a reason for targeted practice, not a reason to add more unrelated reading.

Finally, distinguish official qualification logistics from independent preparation. Anthropic Partner Academy determines eligibility, scheduling, exam policy, scoring and badges. This site’s Anthropic certification listings and supporting articles help readers explore subject matter and choose practice resources. Always book the official exam through the issuer’s current approved process, and keep track of which material is first-party policy versus editorial advice.