Claude Certified Architect – Foundations (CCAR-F): Where the Exam's Weight Sits, and What It Will Not Test
If you are preparing for the Claude Certified Architect – Foundations (CCAR-F) exam, two published facts should decide where your hours go: how its five domains are weighted, and the list of topics the exam says it will not test. Together they tell you what to study hardest and what to leave alone.
What the exam certifies
CCAR-F is not a quiz about a single API call. It certifies that you can architect systems on Claude: design agents that drive their own control flow, run them through real development workflows, keep their context grounded and reliable, prompt them for output you can parse, and extend them with tools and MCP servers that hold up in production. The technologies it covers are Claude Code, the Claude Agent SDK, the Claude API and the Model Context Protocol.
It is built to defeat a vocabulary strategy. Memorising parameter names and definitions does not pass it, because it measures whether you can make an architect's design decisions under realistic constraints, where more than one answer looks plausible and only one holds up. Most questions are scenarios: you are handed a realistic design and asked what to do, or shown a flawed approach and asked to name the failure and the correct alternative. Recall questions exist, but they are the minority.
Where the weight sits
Five weighted domains, largest first:
Agentic Architecture & Orchestration — 27%. The heaviest domain and the heart of the credential: an agent whose control flow is driven by the model's stop_reason, subagents coordinated through a clear router, context passed between agents, and the choice between a fixed pipeline and a dynamic, adaptive workflow.
Claude Code Configuration & Workflows — 20%. Running Claude inside a real development loop — plan, edit, verify — and keeping an autonomous run honest rather than letting it confidently ship something broken.
Prompt Engineering & Structured Output — 20%. Getting machine-parseable output you can act on, reliably, instead of prose you have to regex your way through downstream. The distinction that matters is between asking nicely for JSON and constraining the output so it parses every time.
Tool Design & MCP Integration — 18%. Tools and Model Context Protocol servers whose descriptions and schemas make the model use them correctly rather than guess. It is the smallest domain by concept count, but a fast source of points once the patterns click.
Context Management & Reliability — 15%. Keeping the model grounded in retrieved evidence, making it cite, and recognising where a fluent answer is quietly wrong because its context was thin, stale or unverified. The anti-pattern questions here are the ones learners most often miss.
Scoring is on the total scaled score only — 720 on a 100–1,000 scale — and the per-domain percentages on your result report are informational, not separate bars. So the weights are the plan: an hour spent in a 27% domain buys more than an hour in a 15% one.
What the exam will not test
Anthropic's exam materials list topics that are out of scope. Every hour spent on them is an hour taken from the weighted domains:
- Fine-tuning Claude models or training custom models
- Claude API authentication, billing or account management
- Detailed implementation of specific programming languages or frameworks, beyond what tool and schema configuration needs
- Deploying or hosting MCP servers — infrastructure, networking, container orchestration
- Claude's internal architecture, training process or model weights
- Constitutional AI, RLHF or safety-training methodologies
- Embedding models or vector-database implementation details
- Computer use (browser automation, desktop interaction)
- Vision and image analysis
- Streaming API implementation or server-sent events
- Rate limiting, quotas or API pricing calculations
- OAuth, API key rotation or authentication protocol details
- Specific cloud-provider configurations (AWS, GCP, Azure)
- Performance benchmarking or model comparison metrics
- Prompt-caching implementation details, beyond knowing it exists
- Token-counting algorithms or tokenization specifics
Notice what that list does to a developer's instincts. Streaming, tokenization, authentication and cloud configuration are real engineering work, and none of it is examined here. MCP server design is in scope; MCP server hosting is not.
The scenarios questions are drawn from
CCAR-F is scenario-based. Questions are drawn from Anthropic's official scenarios, and each exam form presents four of them. The official scenarios include Customer Support Resolution Agent, Code Generation with Claude Code, Multi-Agent Research System, Developer Productivity with Claude, Claude Code for CI/CD and Structured Data Extraction. Practise every scenario rather than betting on the draw.
How to study each domain
- Agentic architecture — first and most. Study agentic loops with
stop_reason-driven control flow, coordinator and subagent orchestration, context passing, and the pipeline-versus-workflow trade-off, and pair the reading with a lab where you build a real loop. There is no substitute for having written the control flow yourself. - Claude Code workflows. Practise the scenario questions that ask what a run should do next, not what each command means.
- Context reliability. Over-weight the anti-pattern questions, where a fluent answer is wrong because its context was thin, stale or unverified.
- Prompting and structured outputs. Learn the difference between requesting JSON and constraining output so it parses every time.
- Tool design and MCP. Focus on descriptions and schemas that make the model choose and call tools correctly.
For every concept: read just enough to understand why the mechanism exists, then practise immediately — commit to an answer, rate your confidence, review where your reasoning broke. Interleave the domains rather than cramming one to exhaustion.
What the first domain lesson looks like
The course's first domain lesson — free, with its podcast — is on agentic loops and stopreason control flow, the core of the heaviest domain. It teaches that the loop is model-directed: your runtime sends messages and tools, reads stopreason, executes every requested tool on tooluse, appends the results, and leaves the loop on anything else. Text in a response never decides whether to continue — Claude routinely writes prose and requests a tool in the same message. The API keeps no conversation state, so your runtime resends the whole history every time. Tool results pair with their calls by tooluse_id, not by order, and return together in one user message. And if you know the exact steps in advance, you build a workflow; if you do not, you build an agent.
The mistakes that waste the most time
- Studying out-of-scope engineering — streaming, tokenization, authentication, hosting — because it feels relevant.
- Treating the exam as recall and memorising definitions.
- Over-studying the domain you already find comfortable.
- Reading 720 as 72% and planning by how many questions you can afford to miss.
- Preparing for some scenarios and hoping the draw avoids the others.
How the course maps to the blueprint
Thirteen lessons: an exam-strategy lesson, then twelve domain lessons — three each for Agentic Architecture & Orchestration and Claude Code Configuration & Workflows, and two each for Prompt Engineering & Structured Output, Tool Design & MCP Integration and Context Management & Reliability.
Next step
Take the free CCAR-F diagnostic before you study anything: 24 questions, about 20 minutes, drawn from the same question bank as the rest of the course and weighted like the exam. It reports each domain as questions right out of questions asked — only a few per domain, so a pointer rather than a precise score. The exam-strategy lesson and the lesson on agentic loops are free, with their podcasts.
Alloy is exam preparation built from the official exam guides — only what the exam asks for.