CCAR-F: the architect exam that tests what you normally look up

Claude Certified Architect, Foundations (exam code CCAR-F) is the oldest exam in the program. Until July 2026 it was the only one. It is built around scenarios rather than standalone questions, and it tests exact detail harder than the other three: command options, file paths, parameter values.

The name misleads people. “Foundations” describes the level of detail, not the level of difficulty. It sits below a “Professional” exam that plenty of people find easier, because this one asks you to recall specifics you would normally look up.

Eligibility is the same as the rest of the program: registration runs through the Anthropic Partner Academy with a work email at a Claude Partner Network company. See the program overview.

Exam at a glance

Exam code	CCAR-F
Questions	60. Multiple choice and multiple response.
Format	4 scenarios drawn at random from a set of 6, about 15 questions each
Time limit	120 minutes. Budget about 135 in the chair.
Passing score	720 on a scale of 100 to 1,000
Valid for	12 months
Retakes	14 days, then 30, then 90. Four attempts per rolling 12 months.
Delivery	Proctored. Online with a webcam, or at a Pearson VUE test center.
Prerequisites	None required. The guide suggests six months or more of hands-on building.
Language	English only. No browser translation.

Closed book. The guide describes a solution architect who designs and builds production applications on Claude, with real experience across the Agent SDK, Claude Code, MCP (the Model Context Protocol, the standard way to expose outside tools to a model) and the API. Daily hands-on use matters more here than on any of the other three, because a lot of what is tested is detail you would normally have documentation open for.

The scenario format, and why it changes how you prepare

CCAR-F is the only exam in the program built on scenarios. Six exist. You get four of them, chosen at random, with roughly fifteen questions each. All six are printed in the exam guide, so you can read them before you sit down:

1. A customer support agent built on the Agent SDK with custom tools, a first-contact resolution target, and rules for handing a case to a human
2. Writing code with Claude Code: custom slash commands, CLAUDE.md, plan mode against running directly
3. A multi-agent research system, one lead agent delegating to subagents
4. Developer productivity: exploring unfamiliar code with the built-in tools plus MCP servers
5. Claude Code inside CI/CD: automated review, test writing, pull request comments, cutting false alarms
6. Pulling structured data out of documents: JSON schema validation, odd cases, handing results to other systems

Four out of six means about a third of the possible material will not appear on your sitting, and you cannot know which third. All six need preparing. It also means two people can sit meaningfully different exams, so treat any single “it was easy” or “it was brutal” report as weak evidence.

What it covers, and what each part is worth

Domain	Weight	Roughly
Agentic Architecture and Orchestration	27%	16 questions
Claude Code Configuration and Workflows	20%	12 questions
Prompt Engineering and Structured Output	20%	12 questions
Tool Design and MCP Integration	18%	11 questions
Context Management and Reliability	15%	9 questions

Agent design at 27% is the single largest domain anywhere in the program, and it goes into real specifics. The guide names things like driving an agent loop from stop_reason values rather than reading the reply text for hints that it has finished, the fact that subagents do not inherit the lead agent’s history, launching subagents with the Task tool and the allowedTools entry that requires, issuing several Task calls in one reply so they run concurrently, and the difference between a hook and an instruction: a hook is a guarantee, a prompt is a strong suggestion.

For Claude Code, know the settings precedence: user, then project, then folder-level CLAUDE.md, then imports, rules with path patterns, commands, and skills with their frontmatter options. For structured output, know tool use with JSON schemas, the difference between tool_choice set to auto, any, or a named tool, and how marking fields optional stops the model inventing values.

The appendix nobody opens

The guide’s appendix lists the technology that can appear, which is the closest thing to a syllabus the program publishes. It also lists what is ruled out, and that list is the more valuable half, because it tells you what to stop studying.

Not on this exam: training or fine-tuning models, API billing and account settings, deep language and framework detail, hosting MCP servers, how Claude is built or trained, Constitutional AI and RLHF, embeddings and vector databases, computer use, image analysis, building streaming, rate limits and price arithmetic, OAuth and key rotation, cloud-specific setup, benchmarks, the internals of prompt caching beyond knowing it exists, and token counting.

If a study plan you find online covers half of that list, it was not written from the guide.

Does it count for your employer?

Yes. Architect Foundations is one of the three exams Anthropic’s FAQ lists as counting toward Claude Partner Network eligibility, along with Developer Foundations and Architect Professional. Companies track their certification count in the Partner Portal, updated weekly.

The domain rule applies: register with a work email on a domain recognized on your company’s partner record, or the pass will not be credited to the company.

Preparation checklist

Before you study anything

• Partner Academy account on a recognized work domain
• Read the exam guide end to end
• Read all six scenarios. You are allowed to. Not doing so is leaving free marks behind.
• Read the ruled-out list in the appendix and delete those topics from your plan
• There is no single official prep path for this exam. The relevant courses are separate ones on Claude Code, the API, and MCP.

Hands-on, because this exam rewards muscle memory over reading

• Build an agent loop driven by stop_reason, not by pattern-matching the reply text
• Set up CLAUDE.md at user, project and folder level, then check what actually won
• Write a skill with its frontmatter options and confirm the tool restrictions bite
• Add a hook that blocks a destructive action, and watch it fire
• Run the CLI headless with JSON output and a schema
• Try tool_choice at auto, any, and a named tool on the same request, and see the difference for yourself
• Push one job through the Message Batches API and match results back by their IDs

Practice questions

The official practice exam was retired in the move to Pearson VUE. Anything now advertised as an official practice test is a copy of the retired one. The guide’s sample questions are the only official example of style. Third-party banks are useful for finding weak domains and nothing more.

When it breaks

“I studied prompting and got buried in Claude Code detail.” Predictable from the weights. Claude Code configuration and structured output are 20% each. Prompting alone is not a domain here.

“My scenario set was completely different from my colleague’s.” Expected. Four of six, drawn at random.

“I failed before June 2026 and I am inside a waiting period.” Failed attempts from before the move to Pearson VUE were cleared. You can register again with no wait.

“My registration was rejected” or “my managed laptop blocked the exam”: same causes and same fixes as every exam in the program. They are written out in full on the CCAO-F page.