Claude Certified Architect Foundations: the exam guide
60 questions in 120 minutes across five weighted domains and 4 of 6 published scenarios, 720 to pass, US$125. Here is what it tests and how to prepare in four weeks.
Claude Certified Architect: Foundations (exam code CCAR-F) is Anthropic's exam for solution architects who design and build agent systems with Claude. It is 60 questions in 120 minutes, built around 4 scenarios drawn at random from a published bank of 6, and you need a scaled score of 720 out of 1,000 to pass. The fee is US$125 before any partner discount.
It tests five domains: agentic architecture and orchestration (27%), Claude Code configuration and workflows (20%), prompt engineering and structured output (20%), tool design and MCP integration (18%), and context management and reliability (15%). Every question asks for the best design decision in a production situation, so hands-on building matters more than memorising features. Anthropic's exam guide says the typical candidate has 6 or more months of practical experience with the Claude API, the Agent SDK, Claude Code and MCP.
Only people who work at organisations in the Claude Partner Network can register. If you already build with Claude, four focused weeks on Anthropic's free courses, the exam guide and four practice builds is a realistic plan. If you are a business buying Claude work, the useful question is simpler: who on the team designing your system holds this credential, and is it current?
This post covers the Architect exam only. For all four Claude certifications side by side, plus eligibility and the free Anthropic Academy courses, read our Claude certification guide.
What does the Claude Certified Architect exam test?
Anthropic's exam guide (version 1.0, effective July 2026) splits the exam into five weighted domains and 30 task statements. The weights are the share of scored questions drawn from each domain. The question counts below are our own arithmetic on 60 questions, so treat them as approximate.
| Domain | Weight | About how many questions | What it means in a business build |
|---|---|---|---|
| Agentic architecture and orchestration | 27% | 16 | How an agent decides its next step, splits work between sub-agents, and when code (instead of a prompt) has to enforce a rule |
| Claude Code configuration and workflows | 20% | 12 | Setting up Claude Code for a team: CLAUDE.md files, rules, skills, slash commands, plan mode and CI pipelines |
| Prompt engineering and structured output | 20% | 12 | Getting clean, predictable data out of Claude using JSON schemas, examples and validation with retries |
| Tool design and MCP integration | 18% | 11 | How Claude talks to your systems, such as a CRM or finance tool, through well-described tools and MCP servers |
| Context management and reliability | 15% | 9 | Keeping long conversations accurate, handling errors, and knowing when to hand over to a person |
The six published scenarios are a customer support resolution agent, code generation with Claude Code, a multi-agent research system, developer productivity tools, Claude Code in a CI/CD pipeline, and structured data extraction from documents. You sit four of them, picked at random, so skipping one to save study time is a gamble you cannot control.
From Anthropic's exam guide for CCAR-F, version 1.0, July 2026.
The guide also lists what will not appear. Fine-tuning, vision and image analysis, computer use, streaming, rate limits and API pricing calculations, cloud provider set-up, and tokenisation details are all out of scope. If MCP is new to you, start with what MCP is and how it works before the tool design domain.
What do the Claude architect exam questions look like?
Every item is multiple choice or multiple response, and each one tells you how many answers to select. A scenario sets the scene (a support agent with refund tools, say), then several questions ask what you would change. The sample questions in Anthropic's guide reward two habits: fix the root cause with the smallest change that works, and use code to enforce anything with money or identity attached, because a prompt instruction can still be ignored some of the time.
Below are two questions we wrote ourselves to show the style. They are our own examples, written from the published task statements. They are not exam items and are not taken from any exam.
Our example 1 (prompt engineering and structured output). A finance team uses Claude to extract supplier name, invoice total and ABN from PDF invoices into a JSON schema. About one in ten invoices has no ABN printed, and on those the extraction sometimes returns a plausible but made-up ABN. What is the best first change?
- A. Mark the ABN field as required so the model always looks harder for it.
- B. Make the ABN field nullable and state that it returns null when the number is not in the document.
- C. Add five examples to the prompt, each showing an invoice with an ABN.
- D. Retry the extraction until the ABN field contains 11 digits.
Our answer is B. A required field pushes the model to fill it, which is how invented values appear. Examples that all include an ABN teach the wrong pattern, and retrying cannot recover information that is absent from the source.
Our example 2 (Claude Code configuration). Your team wants Claude Code to follow API naming conventions only when someone edits files under src/api, and to ignore them everywhere else. Where should the conventions go?
- A. In the root CLAUDE.md, with a note saying they apply to API files.
- B. In a file under .claude/rules/ with a paths pattern matching src/api.
- C. In a custom slash command developers run before API work.
- D. In each developer's personal ~/.claude/CLAUDE.md.
Our answer is B. Path-scoped rules load only when matching files are in play. The root file loads for every task, a slash command depends on someone remembering it, and personal files are not shared with the team.
Want Claude built into your business the right way?
We design and build Claude agents that connect to the systems your team already runs.
How much does the Claude Certified Architect exam cost, and how do you book it?
- Fee. US$125 per attempt, set in US dollars and paid by card in Anthropic Partner Academy. It rose from US$99 on 30 June 2026. Any partner tier discount shows at checkout, and so does any tax, so check the total.
- Booking. Register in Anthropic Partner Academy, then follow the link to Pearson VUE to book online proctoring or a test centre.
- Cancelling. Anthropic's certification FAQ says 48 hours before the appointment, while the exam guide says 24. Work to 48. Inside the window, or if you do not show, the fee is gone.
- Retakes. Wait 14 days after a failed attempt, 30 after the second and 90 after the third, with a cap of four attempts in any rolling 12 months. Each attempt costs the full fee.
- Results. Pass or fail with a scaled score from 100 to 1,000, plus your percentage correct in each domain. The domain figures do not decide the pass.
- Validity. 12 months. On-time renewal is a free, non-proctored assessment. If it lapses, you retake the exam at the full fee, and Anthropic may require a full retake anyway if the content changes significantly.
- Confidentiality. You accept a non-disclosure agreement before you start. Sharing exam content can mean losing the credential and a ban from future exams.
Pearson VUE has announced updated content from 7 December 2026 for the Associate, Developer and Architect Professional exams. The Architect exam covered here is not on that list, so the July 2026 guide is still the one to study from. Check the certification page before you book in case that changes.
- Attempt 1Full fee. If you miss 720, wait 14 days.
- Attempt 2Full fee again. If you miss, wait 30 days.
- Attempt 3Full fee again. If you miss, wait 90 days.
- Attempt 4The last one allowed in any rolling 12 months.
Retake rules from Anthropic's certification FAQ and exam guide, checked 5 November 2026.
Who in a business should sit it: architect or developer?
Match the exam to the job the person actually does. Anthropic describes the Developer exam as being for engineers building applications with Claude (the API, tool use and agent development), and the Architect exam for solution architects who design and build agent systems.
| Exam (Foundations level) | Fee, US$ | Best fit in a delivery team |
|---|---|---|
| Associate (CCAO-F) | $99 | Consultants, project leads and staff who use Claude every day |
| Developer (CCDV-F) | $125 | Engineers who write the code that calls Claude |
| Architect (CCAR-F) | $125 | The person who decides how the agents, tools, data and hand-offs to people fit together |
On a typical Claude build, the architect owns the design questions this exam tests: which steps an agent takes on its own, which need a person, and how it reads and writes to your systems. The credential tells you most when its holder is the one making those calls on your project.
For most 50 to 250 person businesses, nobody in-house will sit it, because the exams are limited to Claude Partner Network organisations. What you can do is ask the firm building for you who holds it. Our guide to what the Claude Partner Network is explains how the program works.
A four-week study plan using Anthropic's free material
This plan assumes someone already building with Claude, giving it about six to eight hours a week. The weeks follow the domain weights, so the biggest domain gets a full week. The free courses named here are in Anthropic's own prep list for this exam or its public Anthropic Academy catalogue, and the builds come from the preparation exercises in the exam guide.
| Week | Focus | Free material | Build |
|---|---|---|---|
| 1 | Agentic architecture and orchestration (27%) | Building with the Claude API; Introduction to subagents | An agent with 3 or 4 tools, a loop driven by stop_reason, and a hook that blocks a refund above a set amount and hands it to a person |
| 2 | Claude Code configuration and workflows (20%) | Claude Code in Action; Introduction to agent skills | A team set-up: project CLAUDE.md, path-scoped rules, one skill, one MCP server in .mcp.json, and Claude Code running in a CI job |
| 3 | Tool design and MCP (18%), prompt engineering and structured output (20%) | Introduction to Model Context Protocol; Model Context Protocol: Advanced Topics | An extraction pipeline with a JSON schema, nullable fields, a validation and retry loop, and a batch run |
| 4 | Context management and reliability (15%), then revision | The exam guide in full, including its two sample questions and the out-of-scope list | A coordinator with two sub-agents that keeps source attribution and reports gaps when one sub-agent fails |
Read the exam guide end to end in week one and again in week four. Anthropic calls it the authoritative source for scope, and every question is written against its task statements. Claude 101 is also on the prep list if you want a refresher.
What are the common mistakes on the Architect exam?
- Trusting the prompt to enforce a rule. When a step must happen every time, such as verifying a customer before a refund, the guide favours hooks and code. A firmer system prompt is the tempting wrong answer.
- Picking the biggest fix. Questions often ask for the most effective first step. A routing layer or a rebuilt toolset can work, and still be wrong when better tool descriptions would fix the problem.
- Reading without building. The scenarios assume you have watched an agent misbehave and fixed it. Do the four builds.
- Skipping a scenario. Four of six turn up at random. Study all six.
- Studying what is out of scope. Time spent on fine-tuning, vision or pricing maths earns nothing on this exam.
- Hunting for leaked questions. It breaks the non-disclosure agreement you accept on the day, and the penalty is losing the credential.
Why should a business buying Claude work care?
Because the people who hold this credential have shown they know how to design the parts that go wrong in production: what an agent can do without asking, what it does when a system errors, and how it hands over to a person. If a supplier is building agents that touch your customers or your money, those are the parts you want designed by someone who knows them.
Ask three things. Who on the delivery team holds which credential, by name. Whether it is current, since each one lasts 12 months. And whether that person is designing your build or just on the firm's books. A credential belongs to a person and leaves with them, so pair the answer with examples of shipped work. Our guide on how to choose who builds Claude into your HubSpot has the rest of the questions, and Claude Team vs Enterprise covers the plan decision that sits beside the build.
Neighbourhood is a Certified Partner of the Claude Partner Network. Sophie Costello and Beau Leach on our team hold Claude Certified Architect: Foundations, and they design the Claude systems we build for Australian businesses.
Frequently asked questions
How many questions are on the Claude Certified Architect exam? 60 multiple-choice and multiple-response questions in 120 minutes, framed by 4 scenarios drawn at random from a published set of 6.
What is the pass mark for Claude Certified Architect? A scaled score of 720 on a range of 100 to 1,000. Your report also shows your percentage correct in each domain, but only the total score decides the pass.
How much does the Claude Certified Architect exam cost? US$125 per attempt, before any partner tier discount shown at checkout. Retakes cost the full fee again.
What is the difference between the Claude architect and developer exams? The Developer exam is for engineers building applications with the Claude API, tool use and agents. The Architect exam is for solution architects who design agent systems, including orchestration, Claude Code set-up, MCP tools and reliability.
How long should you study for the Claude Certified Architect exam? If you already build with Claude, about four weeks at six to eight hours a week is realistic. Anthropic's guide says the typical candidate has 6 or more months of hands-on experience.
Sources
- Claude Certified Architect (CCAR-F) certification page and exam guide (PDF, version 1.0, July 2026), Anthropic Partner Academy, checked 5 November 2026, for question count, time, scenarios, domain weights, task statements, scoring, scope, renewal and conduct rules.
- FAQ: Certifications, Anthropic Partner Academy, checked 5 November 2026, for the fee, the June 2026 price change, retakes and the 48 hour cancellation rule.
- Who can take a Claude certification exam, Anthropic Partner Academy, checked 5 November 2026, for eligibility.
- Architect exam prep courses, Anthropic Partner Academy, checked 5 November 2026, for the free course list.
- Anthropic Academy, Anthropic, checked 5 November 2026, for the public course catalogue.
- Four role-based certifications, Anthropic, checked 5 November 2026, for who the Developer and Architect exams are for.
- Claude Certification Program, Pearson VUE, checked 5 November 2026, for delivery options and the December 2026 content update.
See how we design and build Claude agents on our AI engineering page, or get in touch. Tell Matt what you want Claude to do in your business and who would be building it. He will talk you through how we would design it and who on our team would do the work.
Planning a Claude build?
Tell Matt what you want Claude to handle, and he will walk you through how we would build it.