1 The CCAR-F exam · CCAR-F

CCAR-F tests whether you can pick the right architectural fix for a Claude system that is misbehaving in production. It is not a memory test of API fields; it is a judgment test, and most wrong answers are plausible fixes aimed at the wrong root cause.

Where it sits in the program

Anthropic's Claude Certification Program has four role-based credentials, all delivered by Pearson VUE to Claude Partner Network members.

Credential Code Audience Fee

Claude Certified Associate – Foundations CCAO-F Business and knowledge workers using Claude daily US$99 Claude Certified Developer – Foundations CCDV-F Engineers building apps on the Claude API US$125 Claude Certified Architect – Foundations CCAR-F Solution architects designing agentic systems US$125 Claude Certified Architect – Professional CCAR-P Senior architects, enterprise design and governance US$175 There are no formal prerequisites. Anthropic recommends starting the Architect track at Foundations, but you can sit Professional without it, and Foundations does not convert to Professional.

What the exam covers

The guide lists four technologies: Claude Code, the Claude Agent SDK, the Claude API and the Model Context Protocol. Every question hangs off one of six production scenarios, and you get four of them at random.

Figure 1. The five CCAR-F domains by exam weight

Scenario What you are building
Primary do‐
mains

1. Customer Support

Resolution Agent

Agent SDK agent with MCP tools get_customer, lookup_order, pro‐

cess_refund, escalate_to_human; target 80%+ first-contact resolution 1, 2, 5

2. Code Generation with

Claude Code

Team use of Claude Code with slash commands, CLAUDE.md, plan mode 3, 5

3. Multi-Agent Research

System

Coordinator plus search, analysis, synthesis and report subagents producing cited reports

1, 2, 5

4. Developer Productivity

with Claude

Agent SDK tool using Read, Write, Bash, Grep, Glob and MCP servers to explore codebases

2, 3, 1

5. Claude Code for Con‐

tinuous Integration

Automated PR review and test generation with few false positives 3, 4

6. Structured Data Ex‐

traction

Extraction from unstructured documents, validated against JSON schemas 4, 5 The five domains and their weights, which also give you a rough item count out of 60:

Domain Weight About how many items Chapter

1. Agentic Architecture & Orchestration 27% 16 2
2. Tool Design & MCP Integration 18% 11 3
3. Claude Code Configuration & Workflows 20% 12 4
4. Prompt Engineering & Structured Output 20% 12 5
5. Context Management & Reliability 15% 9 6

What is out of scope

The guide explicitly excludes these, so do not spend study time on them: fine-tuning and model training; API authentication, billing and account management; hosting MCP servers; Claude's internals and safety training methods; embeddings and vector databases; computer use; vision; streaming and server-sent events; rate limits, quotas and pricing calculations; OAuth and key

1. Agentic Architecture & Orchestration27%
3. Claude Code Configuration & Workflows20%
4. Prompt Engineering & Structured

Output

20%

2. Tool Design & MCP Integration18%
5. Context Management & Reliability15%

About 16, 12, 12, 11 and 9 of the 60 items.

The ccar-f exam

rotation; cloud-provider configuration; benchmarks; tokenization; and prompt caching beyond knowing it exists.

How it is scored

The exam is criterion-referenced: you are measured against a fixed standard set by subject matter experts, not against other candidates. Scores are scaled from 100 to 1,000 so that different exam forms are equivalent, and 720 passes. Your report shows the percentage correct per domain, but only the total scaled score decides pass or fail.

Logistics

Register on the Anthropic Partner Academy with your partner work email and pay at checkout (partner tiers get discounts). Create your Pearson VUE account from the confirmation email and book online proctoring or a test centre.

Bring government photo ID whose name matches your registration exactly.

Accept the non-disclosure agreement at the start; declining ends the session with no refund. Expect a closed-book exam: no notes, documentation, AI assistants or second monitors. If you do not pass, wait 14 days for the second attempt, 30 for the third and 90 for the fourth, with at most four attempts in a rolling year. On time renewal after 12 months is a free, nonproctored assessment.

How to read a CCAR-F question

Most items describe a symptom from production logs and ask for the most effective fix. Four habits carry most of the exam:

Find the root cause first. The sample questions reward the option that fixes the stated cause, and punish options that fix a neighbouring problem.

Prefer deterministic controls for must-never-fail rules. When money, identity or compliance is at stake, a hook or programmatic gate beats a stronger prompt. Prefer the proportionate first step. A better tool description or explicit criteria usually beats building a classifier, a routing layer or new infrastructure.

Distrust self-reported confidence and sentiment. The guide treats both as poor proxies for case difficulty.

Check yourself

How many scenarios will you see, and from how large a bank?

Which domain carries the most weight, and roughly how many items is that?

Name three topics the guide puts out of scope.

What decides pass or fail: the per-domain percentages or the total scaled score?

The ccar-f exam