Start here
Five guides assume you already know what an agent is, what Claude Code puts in which file, and what a tool contract looks like. This page is where those assumptions get earned — and where the biggest surprise about this exam gets named early: it is more technical than its title suggests.
One honest warning before anything else. Claude Certified Architect — Foundations (CCAR-F) has the most misleading name of Anthropic’s three certifications. “Architect” suggests whiteboards and trade-off talks; the exam will absolutely ask you to reason about designs — but it will also ask you which flag makes Claude Code run headless, what a hook can block, and which context-management tool clears old tool results. In the sitting behind this guide, roughly 25 of the 60 questions were technical in wording, and their wrong answers were harder to eliminate than anything in the practice sets. Anthropic hid the hard questions in the “foundations” exam.
That is not a complaint — it is the study plan. This guide treats the exam as what it is: a technical exam about design. Every page teaches the concepts from the ground up, then collects the exact numbers, names and file paths the exam quotes, then drills the traps — the wrong answers that sound sensible.
01What the exam is
CCAR-F is the entry point to Anthropic’s architect track: 60 multiple-choice and multiple-response questions in 120 minutes, $125, delivered through Pearson VUE online (OnVUE) or at a test centre. You need a scaled 720 out of 1,000 to pass, the credential lasts 12 months, and you get a digital badge through Credly. The scenarios are drawn at random — four of six scenario sets per sitting — which is why two candidates can walk out with honest but opposite stories about what the exam emphasised.
The official page describes the scope as foundational knowledge across four surfaces: Claude Code, the Claude Agent SDK, the Claude API, and Model Context Protocol (MCP). Where the developer exam checks that you can use the API, this one checks that you can design a system around it — and “design” at exam difficulty means knowing what each surface actually does, down to its file names and defaults.
The agentic coding tool that runs in your terminal. Its configuration is a filesystem: CLAUDE.md files for instructions, settings.json for enforcement, hooks, slash commands, skills, subagents. Domain 3 is this surface.
The harness (formerly the Claude Code SDK) that wraps Claude in an agent loop: tools, permissions, sessions, subagents, hooks, MCP. You write code against it; it does the looping. Domain 1 is largely this surface.
The raw Messages API: models, prompts, structured output, caching, batches, context management. Domains 4 and 5 lean on it.
Model Context Protocol — the open standard for connecting tools, resources and prompts from servers to a host like Claude. Domain 2 half of it.
02The five domains — study where the marks are
| Domain | Weight | What it tests |
|---|---|---|
| 01 Agentic architecture and orchestration | 27% | The agent loop, when an agent beats a workflow, coordinator–subagent systems, task decomposition, session state. The heart of the exam. |
| 02 Tool design and MCP integration | 18% | What makes a good tool definition, structured error responses, distributing tools across agents, MCP servers and transports, the built-in tools. |
| 03 Claude Code configuration and workflows | 20% | CLAUDE.md hierarchy, slash commands, skills, path-specific rules, plan mode, iterative refinement, CI/CD. |
| 04 Prompt engineering and structured output | 20% | Explicit criteria, few-shot, schema-enforced output, validation and retry loops, batching, multi-pass review. |
| 05 Context management and reliability | 15% | Long-session context, escalation patterns, error propagation, codebase exploration, human review, provenance. |
Read the table twice, once for the weights and once for a warning. Agentic architecture alone is more than a quarter of the exam; add Claude Code configuration and prompt engineering and three domains carry two thirds of the marks. The warning: every domain ends in specifics you would normally just look up — flag names, parameter values, file locations. The exam does not test that you know where to look things up. It tests that you already looked.
03How it differs from the developer exam
If you came through CCDV-F, calibrate like this. The developer exam asks “what does this feature do?” The architect exam asks “when should this feature be used, and what breaks if you use it wrong?” Same documentation, different verb. A developer question names a tool and asks for its purpose; an architect question hands you a half-designed system and asks which piece is missing or misplaced.
Three concrete shifts follow. First, multi-agent thinking moves to the front: CCDV-F meets subagents late and briefly; here they are a quarter of the exam, with coordination, context passing and failure handling in the open. Second, Claude Code deepens from one subdomain to a full domain — configuration files, precedence rules and headless automation. Third, reliability becomes a design subject of its own: escalation ladders, error propagation, confidence calibration, provenance. That is the architect’s half of the job — the half nobody outsources to the model.
04The map from here
The five guides that follow are self-contained; each teaches its domain from the ground up, then pins the numbers, then sets the traps. If you study in weight order, the sequence is Domain 01, then 03 and 04 (tied at 20%), then 02, then 05. If you are coming straight from the developer track, Domain 01 first is still right — it is the largest and the least familiar. The hub’s study-order toggle sorts all of this for you.
Study pointers
- Official CCAR-F page — registration, exam guide and policies ↗
- The CCDV-F study guide — the developer-track floor this guide assumes
- Claude Certified Architect Foundations (CCAR-F) exam prep — the experience write-up behind this guide
- Evidence and transparency — every claim in this guide, sorted by evidence class