Claude Certified Architect Foundations (CCAR-F) Exam Guide (2026)

If you can decide when a coordinator should spawn a subagent, when a hook must block a refund, and when plan mode beats a direct Claude Code run, you are in the right exam.

CCAR-F is Anthropic's first technical Claude certification. It is not Associate Foundations. It is not Developer Foundations. It is not Architect Professional. The official exam guide names the audience as a solution architect who designs and implements production applications with Claude. The registration page says the same job in other words. The sitting tests whether you can make informed tradeoffs on real Claude work.

This guide is for people who will sit the current Version 1.0 blueprint, effective July 2026. Weights, task names, and scenario names below are copied from that PDF.

Who this exam is for

Take it if you already build with the Claude Agent SDK, Claude Code, the Claude API, and Model Context Protocol. Anthropic recommends six or more months of that work. The items ask for tradeoffs in production, not a definition of a prompt.

Skip it if you only chat with Claude. The Associate exam is the productivity-tool credential. Skip it if you want enterprise governance, stakeholder programs, and lifecycle management. That is Architect Professional. Skip it if your job is shipping application code against the API with little architecture work. That is Developer Foundations.

You do not need to hold another Claude credential first. Foundations does not convert into Professional. The certification FAQ states that split in those words. You can sit Professional without Foundations. Most people still start here because the Version 1.0 tasks match the work of standing up agents, tools, and Claude Code before you govern a firm-wide program.

The candidate Anthropic wants already feels the constraint in the stem. A refund that must never exceed a cap is a hook. A research job that should not always run the full pipeline is a coordinator. A one-function fix is direct execution. A 14-file design is plan mode. People who pick the product they memorized last week fail those items.

Sit access is partner-gated

You cannot buy a seat with a personal Gmail or Outlook address. Anthropic's certification FAQ says registration needs a Claude Partner Network company email on a recognized domain. Personal email is rejected. Membership applications are free at claude.com/partners. Adding a domain to a partner record takes 7 to 10 days.

If you sign in and see "You aren't signed in with a partner company employee email," you are not ready to schedule. Fix the domain first. Certifications earned on a domain that is not on the partner record do not credit to the firm, including some subsidiary and regional domains.

You must be 18 or older. Age is checked at Pearson against a government ID. The exam is English only and closed book. Browser translation tools are not allowed under Pearson proctoring. Notes, docs, and AI assistants are also out.

Register and pay in Anthropic Partner Academy. Pearson then emails the scheduling link. You can sit online proctored or at a Pearson test center. After you pay, the registration stays valid for 5 years. There is no short window that forces you to sit next week.

Cancel or reschedule at least 48 hours before the appointment. A late change or a no-show forfeits the fee. Cancelling the Pearson slot does not issue a refund by itself. Refunds go through Anthropic certification support. The name on the Pearson profile must match the government ID exactly. A mismatch at the door also forfeits the fee. Ask for a name correction at least 24 hours before the seat.

Exam shape

The official CCAR-F exam guide, Version 1.0, weights five domains.

Domain 1 is more than a quarter of the scored items. If you only memorize CLAUDE.md paths and skip stop_reason and coordinator design, the math is against you.

The sitting is 60 items. Item types are multiple choice and multiple response. Each item states how many answers to select. Seat time is 120 minutes of questions, about 135 minutes with check-in. The exam uses 4 scenarios drawn at random from a published bank of 6.

Official scenario bank from the PDF:

1. Customer Support Resolution Agent
2. Code Generation with Claude Code
3. Multi-Agent Research System
4. Developer Productivity with Claude
5. Claude Code for Continuous Integration
6. Structured Data Extraction

Delivery is Pearson Professional Assessments. You schedule through Pearson VUE after you pay in Anthropic Partner Academy.

The scaled score runs from 100 to 1,000. The pass mark is 720 on every Claude exam, including this one. Scoring is criterion-referenced. You pass the exam as a whole. Domain percents on the score report are for study, not a second pass bar.

The badge lasts 12 months from the award date. Credly by Pearson issues it. On-time renewal is a free open-book Partner Academy assessment. A lapsed credential requires the full exam again.

The list price is $125 USD. The FAQ says the fee was $99 until 30 June 2026. Partner-tier discounts apply at checkout. Registered-tier partners pay full price. Select, Preferred, and Global Premier partners get 50 percent off. Through 31 December 2026, Global Premier partners get 100 percent off, then the 50 percent rule returns. Retakes cost the same fee and follow a longer wait after each fail. 14 days, then 30, then 90. You may sit the same exam four times in 12 months.

CCAR-F counts toward Claude Partner Network tier standing. Associate Foundations does not.

What changed in 2026

Anthropic launched Claude Certified Architect Foundations on 12 March 2026 as the first Claude technical certification, next to the Claude Partner Network. Associate Foundations, Developer Foundations, and Architect Professional arrived later in the four-credential expansion. The Partner Academy hub now lists those four Pearson exams.

On 30 June 2026, delivery moved to Pearson and badges moved to Credly. The official practice exam on the old platform was retired. Sample items now live only inside the exam guide PDF. Early six-month badges were extended to 12 months from their original earned date. If you failed Architect Foundations before that move, the FAQ says the earlier fail was cleared and the Pearson wait does not apply to that attempt.

The exam guide in force is Version 1.0, dated July 2026. Study notes that stop at "write a better prompt" miss hooks, MCP resources, plan mode, and Message Batches.

Academy course-completion badges on academy.claude.com are not this exam. Anthropic says those quizzes are not the Pearson and Credly credentials.

How orchestration choices fail

Domain 1 is not a glossary of agent words. It is a constraint list. The stem will tell you whether a rule must never fail, whether one agent should pick the next worker, and whether the session needs to resume or fork.

Read the stem for the rule that cannot drift before you pick an architecture. Implement the agentic loop on stop_reason. Continue on tool_use. Stop on end_turn. Do not parse the model's prose to decide the loop is done.

Use an Agent SDK hook when a refund cap, an identity check, or another policy must block the tool call. Prompt text that says "never refund over $500" is not a guarantee. The hook intercepts the call.

Use a coordinator that picks a subagent when the next step depends on the last finding. Pass context in the next prompt. Subagents do not inherit parent memory. A crash-safe workflow keeps a structured manifest so a resume does not invent the last state.

Use a single agent loop when the job is one bounded tool path and a coordinator would only add hops. Parallel subagents help when independent research slices can run at once. Sequential handoff helps when each step needs the prior result.

A useful drill: write one sentence per story that names the rule that cannot fail and whether a second agent needs a fresh context. If you cannot name those two facts, you are guessing the logo.

Session resume and fork show up on the same domain. Resume when the work continues from saved state. Fork when you need a branch that must not pollute the parent transcript.

How to study without wasting a month

Week 1. Domain 1 loops. Implement an agentic loop on stop_reason. Design a coordinator that picks subagents instead of always running the full pipeline. Pass findings in the next prompt. Use hooks when a refund cap or identity check must never fail. Practice task decomposition until you can say why the work splits, and practice handoff until you can say what the next agent is allowed to see.

Week 2. Tools, MCP, and Claude Code. Write tool descriptions that name inputs, outputs, and the neighbor tool you should not pick. Return structured MCP errors with retry metadata. Give each agent a short tool list. Configure project MCP in .mcp.json and personal MCP in ~/.claude.json. Put team rules in project CLAUDE.md, not in ~/.claude/CLAUDE.md. New hires never see a user-level file.

Know plan mode for multi-file design and direct execution for a one-function fix. In CI, use non-interactive -p and JSON output so the pipeline can score the result. Custom slash commands and skills need a scope. Project scope is the team default. User scope is a personal shortcut. Path-specific rules in .claude/rules/ load conventions only where the glob matches.

Week 3. Prompts that a machine can score. Replace "be conservative" with criteria a reviewer can apply twice. Use JSON Schema and tool_use for structured extraction. Add few-shot examples for the ambiguous cases, not for the easy ones. Use Message Batches when a 24-hour window and no latency SLA are acceptable. Keep a real-time call when a developer is blocked on the result. Retry and validate before you invent a second model. Multi-pass review is for fields that fail a confidence bar, not for every row.

Week 4. Context, reliability, and the six scenarios. Budget tokens. Summarize as you go. Trim verbose tool output. Keep claim-to-source maps. Decide when a human must see the output. For each official scenario, write one sentence that names the constraint. First-contact resolution versus escalate. Plan mode versus a slash command. Coordinator versus a single agent. Built-in tools versus an MCP server. CI review that must not spam. Extraction that must not invent a field.

Read the sample items in the official PDF for format. They are the official practice items after the Pearson move.

Do not spend the month fine-tuning a model "for the cert." Fine-tuning is out of scope. Spend the month writing the constraint, then picking the control. That is the exam habit.

Traps that look like easy elimination

Easy banks teach the wrong habit. Real CCAR-F items put two good Claude designs in the list and ask which one matches the constraint.

• Fine-tuning, embeddings, vision, computer use, and cloud-provider setup are out of scope on the official guide. If the stem wants a new trained model, you are practicing the wrong exam.
• Hosting an MCP server is out of scope. Designing the tool and resource interface is in scope. Resources catalog content. Tools take actions.
• A larger context window does not fix attention that thins across a 14-file review. Split the work.
• User-level CLAUDE.md is not a team standard. New hires will not see it.
• Giving a synthesis agent every tool you have makes selection worse. Four or five scoped tools beat a pile.
• A prompt that says "never refund over $500" is not the same as a hook that blocks the tool call.
• Academy quiz badges are not Pearson credentials and do not count toward partner tier.
• A personal email that can open Claude.ai still cannot register for this exam.
• API keys, billing, rate-limit math, OAuth details, and prompt-caching internals are out of scope. Knowing that prompt caching exists is enough.
• Streaming, token-count algorithms, and model-weight trivia do not appear as jobs this candidate is expected to do.
• If you can delete the scenario and still pick the answer from a product name alone, the question is too easy. The live exam will not do that.

How this maps to CloudFluently

Study with the Claude Certified Architect Foundations practice exam sets. Use this page for how to study and why the official domains matter. Use Anthropic for task statements, weights, and sit rules.

Official exam guide: Claude Certified Architect Foundations exam guide PDF. Official prep: Architect Foundations prep courses. Register from Architect Foundations certification after your partner domain works. Program rules live on the certification FAQ and the certification policies pages.

Frequently Asked Questions

Is CCAR-F the same exam as Associate Foundations? No. Associate Foundations is the productivity-tool credential and does not count toward partner tier. CCAR-F is the architect sitting for production applications with the Agent SDK, Claude Code, the API, and MCP.

Do you need another Claude certification first? No. Foundations does not convert into Professional. You can sit Professional without it. Anthropic still calls Foundations the natural start.

Can you register with a personal email? No. Registration needs a Claude Partner Network company email on a recognized domain. Personal email is rejected. Domain adds take 7 to 10 days.

How many questions are there, and how long is the seat? Sixty items. 120 minutes of questions. About 135 minutes with check-in. Four of the six official scenarios appear on a sitting.

What is the pass mark? A scaled 720 on a 100 to 1,000 scale. You pass the exam as a whole. Domain percents on the score report are study hints.

How long does the badge last? Twelve months from the award date. On-time renewal is a free open-book Partner Academy assessment. A lapsed credential requires the full exam again.

What should you study first? Agentic loops, hooks, and coordinator design. Domain 1 is 27 percent. Then tools and MCP, then plan mode versus a direct Claude Code run.

Where is the official outline? Claude Certified Architect Foundations exam guide PDF. Register and book from Architect Foundations certification. Use Anthropic for task statements and weights. Use this page for how to study and what the 2026 Pearson move changed.