An organization can demonstrate Claude answering a question in minutes and still struggle to put that capability into production. Production work introduces tools, permission boundaries, unreliable data, latency budgets, human approval, incident handling, and a question that a polished demonstration cannot answer: what happens when the model makes a plausible but consequential mistake? Claude Certified Architect – Foundations is concerned with those decisions, not merely with writing effective prompts.
The credential is issued by Anthropic. Its published exam guide identifies CCAR-F as the formal exam code; CCA-F is a widely used shorthand and appears in the existing ExamTopics.info exam URL. Those labels refer to the same Architect – Foundations certification, not two editions of it. Candidates comparing booking information should use the issuer’s current wording rather than assuming every third-party abbreviation is an official program code.
What the certification actually measures
An architect’s task is to choose where reasoning belongs and where software must retain control. In a conventional application, a workflow may be a series of defined steps with predictable branching. An agentic application can select tools and revise its plan as evidence arrives, but that flexibility introduces additional ways to fail. The right design may use an agent for interpreting ambiguous requests and conventional code for authorization, state transitions, financial calculations, or irreversible actions.
The exam therefore looks at a connected system: the Claude API, Claude Agent SDK, Claude Code and Model Context Protocol (MCP). A strong candidate understands the boundaries between those components. Claude Code is not synonymous with the Agent SDK; an MCP server is not an authorization policy; a JSON-shaped answer is not proof that its contents are correct. Those distinctions repeatedly affect architectural judgment.
Anthropic certifications serve different professional roles. Claude Certified Associate – Foundations emphasizes broad Claude literacy and responsible professional use.
Claude Certified Developer – Foundations addresses building Claude applications and integrations, while Claude Certified Architect – Professional moves further into enterprise solution design and governance. CCA-F assesses architectural choices in building production Claude systems; the credentials represent different work rather than a compulsory ladder.
The published exam blueprint
Anthropic Academy’s current public listing describes a 60-item, 120-minute exam with multiple-choice and multiple-response questions. It lists a scaled passing score of 720 on a 100–1,000 scale, a US$125 fee, a 12-month credential validity period, and online-proctored or test-center delivery. The formal exam guide says four scenarios are drawn from a bank of six. These are issuer-published details at the time of research, not guarantees about a future registration date; candidates should review Anthropic’s live exam listing and policies before paying or scheduling.
The five content domains provide a more useful study plan than a generic list of Claude features:
- Agentic Architecture & Orchestration — 27%. Choosing between agents and deterministic workflows, handling tool calls, delegation, session boundaries and escalation.
- Tool Design & MCP Integration — 18%. Designing safe tool interfaces, choosing clear contracts and keeping access controls outside the model’s discretion.
- Claude Code Configuration & Workflows — 20%. Project instructions, permissions, hooks, review loops and controlled developer automation.
- Prompt Engineering & Structured Output — 20%. Reliable instructions, examples, extraction, schema requirements and validation.
- Context Management & Reliability — 15%. Maintaining useful state under context limits, preserving evidence and recovering from partial failure.
The weights should influence the attention given to each domain, but they do not mean an architect can ignore the smallest one. If a case asks for a resumable research workflow, context management may determine whether the proposed agent is fit for purpose even when the question appears to concern orchestration.
What candidates can actually register for
The public certification listing describes the exam, but booking eligibility is a separate control. Anthropic’s current certification FAQ restricts registration to Claude Partner Network organizations, requires an eligible corporate email domain and sets a minimum candidate age of 18. It also says there is no mandatory earlier certification. An architect can therefore be technically prepared without being currently eligible to register. This distinction matters to readers planning a budget, a partner training program or a timeline for an individual certification attempt.
The exam is scenario-framed: four scenarios are selected from a disclosed bank of six. They include a customer-support resolution agent, a Claude Code development task, a multi-agent research system, developer tooling, Claude Code in CI and structured extraction. A useful way to study is to identify the control decision in each setting. For a support agent, compare a text-only recommendation with a state-changing tool call; for a research system, ask what source evidence survives delegation. For a CI task, distinguish an assistant’s assertion from an external test result. Practicing those boundaries is more useful than trying to predict which four contexts you will see.
The issuer’s CCAR-F code and ExamTopics.info’s existing CCA-F abbreviation identify the same credential. By contrast, Claude Certified Associate – Foundations and Claude Certified Architect – Professional are different exams with different purposes; the Professional qualification is not awarded automatically when someone passes Foundations. Readers deciding between credentials should rely on Anthropic’s live certification eligibility and policies, since booking rules, partner discounts and exam delivery can change independently of the technical objectives.
Architecture decisions the exam is likely to make you compare
Consider a customer-support assistant that receives an account cancellation request. One proposal allows the model to look up the account and execute a cancellation immediately. A second proposal allows it to identify the request, request the necessary account evidence, and hand the irreversible operation to a service that checks authorization and records the outcome. The second design contains more safeguards because the action’s consequences exceed the confidence we can place in text alone.
Now change the scenario: a research assistant must compare twenty documents, extract claims and identify contradictions. A rigid rule engine may be too brittle for finding thematic differences across unfamiliar documents. An agent can explore, ask targeted retrieval questions and synthesize tentative findings, provided that sources are preserved and a human or validation layer checks unsupported claims. The best boundary is not “agent everywhere” or “never trust an agent.” It is a division of labor that matches uncertainty to risk.
The same reasoning governs deployment platform choices. Direct Claude API integration may suit a team that wants explicit control over service orchestration and observability. A cloud-managed integration may fit an organization whose networking, identity, logging and procurement already live in a particular cloud. The choice depends on system requirements, not on the platform’s name. In agentic applications using Amazon Bedrock, orchestration similarly depends on tool boundaries, delegated tasks and failure handling, although the platform’s identity and deployment controls differ.
From prompt to tool to application boundary
A prompt can tell a model to avoid unsafe operations, but an instruction is not an enforcement mechanism. A system must validate who can request an action, what data the tool may return, whether the action needs confirmation, and which durable system records the result. Effective tools expose narrowly defined operations such as lookup_order or prepare_refund_request instead of an unrestricted database executor. They return typed, bounded results and error states that the application can interpret reliably.
MCP standardizes ways for an application to describe and access tools, resources and prompts. It is useful for integration, yet it does not automatically authenticate users or make an untrusted result safe. Architects should be able to trace data from external systems through tool responses into Claude’s next decision, identify where access controls apply, and describe what evidence would be logged if the tool response proved wrong.
Structured output solves a different layer of the problem. A validated response schema can prevent a malformed object from flowing into a downstream service. It does not guarantee that the extracted invoice number or the proposed security classification is correct. The content still needs checks against source evidence and business rules. A production design often combines schema validation, semantic verification and human review rather than expecting one mechanism to replace the others.
Claude Code belongs in the same architectural picture
Claude Code can assist with changing a repository, investigating tests and preparing reviews. But access to a local development environment is not permission to alter production. A team needs a clear source of project instructions, an explicit permission model, meaningful checkpoints, and verification performed by tools other than the model’s own description of its work.
Repository-specific instructions in CLAUDE.md can help a coding agent understand conventions and architectural constraints. Hooks can call deterministic checks at defined workflow points. Skills can supply scoped procedures. None should be mistaken for a substitute for protected branches, separate deployment credentials, code review or automated tests. A design review should ask how a proposed coding workflow behaves when a test fails, a user interrupts execution, or a suggested command would affect sensitive files.
Agent SDK applications and Claude Code workflows may share concepts—tools, context, delegation and audit—but they need not share the same security posture. A production-facing agent that can update customer records and a coding assistant allowed to edit a local test repository have different owners, risk profiles and approval rules.
Reading scenario questions as an architect
The exam guide’s scenario framing is an invitation to think in constraints. Before choosing an answer, identify the system’s non-negotiable properties: data boundaries, consequences of failure, latency, cost, need for freshness, audit requirements, and whether a human can intervene. Determine which parts require probabilistic interpretation and which require deterministic guarantees. Then ask whether the proposed design supplies evidence that those guarantees hold.
For example, if a recommendation says “instruct the model never to call the delete tool without approval,” it may sound responsible. But if deletion is genuinely high risk, a stronger design enforces approval in application logic and gives Claude no route around it. Conversely, building a complex external rules engine to distinguish two clearly described document types may add avoidable complexity if a clearer prompt and a few examples would produce an adequate first-pass classification subject to validation.
Use this principle when comparing orchestration patterns. A subagent earns its place when isolated context, parallel research or a different tool set improves the task enough to offset coordination cost. It is not necessary just because multiple topics appear in a prompt. A tool result should preserve enough provenance for review, but it need not dump an entire database record into the model’s context if the next decision uses only a few fields.
Preparing with evidence rather than memorized recipes
A useful preparation exercise begins with a small architecture decision and ends with an observable artifact. Design a research workflow that retrieves documents, records citations, synthesizes a result and surfaces unresolved conflicts. Write down what terminates the agent loop. Specify the maximum tool attempts and the recovery procedure if a service times out. Add a test in which a retrieved document contains malicious instructions and verify that it is treated as data rather than authority.
For the Claude Code domain, create a disposable repository with project instructions, narrow permissions and a test hook. Make a change, inspect the generated diff and verify the test runner’s result. For structured output, try an extraction prompt against messy documents with missing fields and unexpected line breaks, then compare model output with independent schema and semantic checks. For context management, interrupt a multistep task and see which facts must be preserved to resume safely.
Do not infer a raw percentage from the exam’s scaled passing score. A score of 720 is not a statement that 72 percent of items must be correct. Focus on choosing proportionate controls and explaining the tradeoff behind each design. This creates transferable architectural competence and prepares candidates for variations that no practice-question bank can predict.
The next architectural decision is usually not how to make Claude say more. It is which facts it needs, what authority it should possess, what software can verify, and how the surrounding system will recover when an assumption fails. Those are the decisions at the heart of Claude Certified Architect – Foundations.