The person who can make Claude produce an impressive demonstration is not necessarily ready to design the system that a customer will operate. Anthropic’s Claude Certified Architect – Foundations exam asks candidates to weigh failure modes, tool boundaries, deployment choices and evidence in realistic application scenarios. Understanding how the assessment is organized is useful because a scenario question often offers several technically possible answers; the better choice is the one that fits the stated operational constraints.
The certification is often shortened to CCA-F. Anthropic’s published exam guide uses CCAR-F as the formal examination code. These are two labels for the same Architect – Foundations credential. They are not the Associate – Foundations (CCAO-F), Developer – Foundations (CCDV-F), or Architect – Professional (CCAR-P) qualifications. Recognizing an agent’s permission and evidence boundaries is part of the technical preparation; candidates also need to understand exam mechanics, eligibility, and scenario-based assessment.
Check eligibility before choosing an exam date
Anthropic’s certification program is currently restricted to people at Claude Partner Network organizations. Its certification FAQ says that registration requires a recognized work email address associated with a partner organization and that individuals outside the network cannot currently register. An interested developer’s skill level alone does not confer eligibility. Anthropic also requires candidates to be at least 18 and to meet identity-verification requirements. If a training provider describes this as an unrestricted public exam, check its statement against Anthropic’s current partner eligibility and registration rules before paying for anything.
There is no mandatory earlier Claude credential: Architect – Foundations does not require passing Associate – Foundations or Developer – Foundations first. That is a certification prerequisite question, different from the practical experience needed to answer architecture scenarios reliably. Someone who has never dealt with a failed tool call, a resumable job, or a privileged operation can memorize names yet struggle to justify a design decision. Build at least one small project whose logs and failures you can inspect before relying solely on reading material.
Partner access, discounts, geographic exam-delivery rules and exam policies may change. The live issuer listing, rather than a historical blog statement, controls who can book and how an exam is administered. Your company’s training coordinator can help verify recognized organizational access, but that is not the same thing as certifying an individual automatically.
The current examination contract
As listed by Anthropic’s Architect – Foundations certification page, the exam contains 60 items with 120 minutes of answering time; total seat time is approximately 135 minutes after check-in and related steps. Multiple-choice and multiple-response questions appear. A multiple-response question indicates how many options to select. Never treat every item as single-answer merely because older practice materials used that format.
The passing mark is 720 on a scaled 100–1,000 reporting scale. It does not mean 72% of questions correct, nor is there a published safe number of mistakes. Scaled scores account for differences between forms; use domain feedback to improve study decisions instead of reverse-engineering a percentage cutoff. The issuer lists a US$125 standard fee and 12-month credential validity, subject to its current registration terms and organizational discounts. At the time of review, delivery is online proctored or at a Pearson VUE test center.
Anthropic’s FAQ describes a registration as valid for five years, but a purchased registration should not be confused with the life of a certification after passing. It also specifies progressive retake waiting periods: 14 days after the first unsuccessful attempt, 30 after the second, and 90 after the third, with at most four attempts per exam in a rolling twelve months. Retakes require a new registration and fee. Verify these operational policies shortly before your actual appointment because the provider, market rules and support arrangements can change independently of the technical blueprint.
Read the five weights without inventing a question count
The published Architect – Foundations blueprint places 27% on Agentic Architecture & Orchestration, 18% on Tool Design & MCP Integration, 20% on Claude Code Configuration & Workflows, 20% on Prompt Engineering & Structured Output and 15% on Context Management & Reliability. These percentages describe blueprint weighting, not guaranteed exact counts by topic on your particular exam form. It would be misleading to promise a fixed number of MCP or context questions.
Each domain contains decisions rather than isolated terminology. In orchestration, you should be able to decide whether to invoke a model, run a deterministic step or delegate work to another agent. Tool design asks where authorization and input validation live. Claude Code asks how repository instructions, permissions, hooks and CI evidence affect developer workflows. Structured output asks what a schema can enforce and what source-grounded validation must still check. Reliability asks what survives a restart and when an incomplete answer must be escalated. A candidate who learns these distinctions can apply them when unfamiliar details change.
Allocate study time broadly in proportion to these weights, but do not abandon the smallest domain. For instance, a customer-support agent question may appear principally about tool permissions while the decisive detail is whether a failed account-update request has a safe retry procedure. In an agent loop with delegation, a failed state-changing operation must be reconciled before a retry even when orchestration is the primary question.
How the four-of-six scenario structure changes preparation
The formal guide describes four scenarios drawn from a published bank of six. The scenarios are contextual frames for several questions, not instructions to memorize a single correct architecture. You do not know which four you will encounter. Train on the possible decisions and on the facts that constrain them, not on guessing the chosen subset.
Customer Support Resolution Agent. A support assistant must interpret an ambiguous request, retrieve customer information and possibly trigger a consequential operation. The interesting question is not whether Claude can phrase a refund politely; it is where identity checks, approval, idempotency and actual success verification occur. Sketch the control boundary before choosing a tool design.
Code Generation with Claude Code. A coding task must obey repository instructions, edit the correct files and produce trustworthy test evidence. Consider which rules are durable project instructions, which scripts are deterministic release gates and how a failed test should affect the reported outcome. Do not confuse a fluent explanation with an executed test result.
Multi-Agent Research System. Parallel research can isolate context and shorten some workloads, but it introduces duplicated retrieval, inconsistent sources, handoff errors and incomplete synthesis. Practice deciding which work truly benefits from subagents and what provenance the coordinator must carry forward. A final answer should surface missing source coverage, not mask it.
Developer Productivity Tooling. Tools, user approvals and environment configuration influence whether a coding assistant can carry out a safe task. Ask when a built-in tool is sufficient, when a narrowly scoped MCP tool earns its place and whether permissions differ between local development and shared environments. A narrowly scoped MCP tool contract should identify permitted operations, required identity, validated inputs and the meaning of a rejected request.
Claude Code in a CI Pipeline. A command that completes normally is not proof that a change is suitable to merge. The workflow needs predictable input context, protected credentials, actionable review output, exit-state handling and independent tests. A missing or failed test must appear as a failure, not as a model-generated claim of success. In a Claude Code CI review, completion of the assistant’s analysis does not turn a failed test into a pass or override branch protection.
Structured Data Extraction. The model may return a valid object even when source facts are wrong, absent or conflicting. Decide which fields may be null, how evidence is recorded, when a batch entry is retried and when it requires review. A reliable batch-extraction architecture validates output shape separately from evidence that each extracted field is supported by the correct source.
A decision method for multiple-response questions
Start by translating the scenario into an explicit contract: who requests the action, what changes if it succeeds, which data may be trusted, what must happen when a tool fails and what evidence proves completion. Then compare the options against those constraints. An answer that sounds secure because it says “tell the model not to” is weaker than one that enforces the same boundary in the service executing the action. An answer proposing an additional agent can also be worse if the task needs only a deterministic lookup and adds unnecessary uncertainty.
For a select-two item, test each chosen option independently against the requirements and then test the pair. Two apparently attractive recommendations can conflict: one may permit a direct privileged action while the other requires a human to approve every such action. Multiple-response preparation therefore needs careful reading and explicit reasoning, not just a list of fashionable technologies. Rejected options are part of your study evidence: write down which requirement they violate.
A useful practice technique is to change one variable in a scenario. A support assistant that only drafts a proposed address change can rely on different protections from one allowed to submit the change. A research assistant that reports tentative findings does not need the same authorization as one exporting customer records. If your preferred architecture never changes when the consequence changes, you are probably memorizing a pattern rather than applying architectural judgment.
A four-week preparation sequence with observable work
Week one: agent decisions and tool boundaries. Build a small support workflow that can retrieve a record and propose—but not execute—a sensitive change. Implement a server-side permission test and record the decision, not just the generated message. Simulate an unavailable service and a response lost after an operation. Produce a diagram showing which component owns each policy. Study the real differences between text, tool requests, and tool results.
Week two: Claude Code and MCP. Put a practice repository under simple conventions, decide which instructions belong in project memory and which checks are executable. Configure a narrow tool with explicit arguments, error handling and least privilege. Test one requested action under three identities: authorized, unauthorized and unknown. Explore the distinction between reusable instructions and runtime permissions.
Week three: structured output and reliability. Extract a small set of fields from heterogeneous documents. Include contradictory statements, missing values and malicious text attempting to redirect the model. Validate structure separately from source evidence; then interrupt and resume the work. Record a durable state checkpoint and compare actual source citations after recovery. Add a failure that should produce human escalation rather than a fabricated completion.
Week four: scenario simulation. Create a timed set containing questions from all five domains and all six scenario types. Record the reasons for correct and rejected design choices, revisit weak constraints and repeat the exercises with changed assumptions. Set a realistic time budget: two minutes per item is the mathematical average across 60 items and 120 minutes, not a recommendation to spend exactly two minutes on every item. Leave review time for multi-response questions and ambiguous wording. Study the issuer’s sample-item format rather than seeking confidential live questions.
Evidence that you are prepared—and what remains outside your control
A credible readiness package includes an agent-loop diagram, a tool contract with access controls, a small repository’s instructions and CI results, a schema-and-evidence extraction example and a restart test. Someone reviewing those artifacts should be able to tell whether the workflow performed an operation, merely proposed one, or failed before completing it. If that distinction is absent, more practice-question volume will not cure the underlying weakness.
Before scheduling, check the issuer’s active eligibility, fee, test-center or remote testing requirements and current exam guide. A proctored exam has operational rules independent of technical preparation, including accepted ID and workstation restrictions. None of these exercises promises a passing score. They are valuable because they exercise the architectural judgments Anthropic says the credential represents.