CCA Exam Domains Explained: All 5 CCA-F Domains and Weights

The Claude Certified Architect – Foundations (CCA-F, exam code CCAR-F) exam covers five domains: Agentic Architecture & Orchestration (27%), Tool Design & MCP Integration (18%), Claude Code Configuration & Workflows (20%), Prompt Engineering & Structured Output (20%), and Context Management & Reliability (15%). Every one of the 60 scenario questions maps to one of these domains, and the weights decide roughly how many questions you will see from each.

This post walks through what each domain actually tests, which is usually narrower and more opinionated than the domain title suggests. Weights are as of mid-2026; confirm on the official Anthropic / Pearson VUE page before you register.

How are the 5 CCA exam domains weighted?

Domain	Name	Weight	Approx. questions of 60
D1	Agentic Architecture & Orchestration	27%	~16
D2	Tool Design & MCP Integration	18%	~11
D3	Claude Code Configuration & Workflows	20%	~12
D4	Prompt Engineering & Structured Output	20%	~12
D5	Context Management & Reliability	15%	~9

The question counts are arithmetic on the percentages, not a published guarantee. Two things stand out: D1 alone is more than a quarter of the exam, and no domain is small enough to skip. If you scored zero on D5 you would need to be almost perfect everywhere else to reach the 720 scaled passing score.

Our study guide uses this numbering everywhere. Note that some other sites order the domains differently; the content is the same, the labels are not.

What does Domain 1 (Agentic Architecture & Orchestration) really test?

D1 is about how an agent loop runs and how multiple agents are coordinated. The exam is less interested in “what is an agent” and more in “you have this failure, which design fixes it”.

Representative topics:

• The loop and its stop signal. Questions expect you to know that stop_reason drives the loop, that the API is stateless, and that hardcoded decision trees are the wrong tool when the model should be routing. See stop_reason Is the Only Loop Control That Matters and Model-Driven Decisions vs Hardcoded Decision Trees.
• Hub-and-spoke orchestration and sub-agent isolation. Sub-agents start with an empty context; anything they need must be passed in explicitly. Sub-Agents See Nothing is the single most-referenced idea in this domain.
• Programmatic vs prompt enforcement, and hooks. If a rule must hold 100% of the time, a prompt is the wrong place for it. The Model Will Override Your Instructions 4-15% of the Time and the hooks articles under Task 1.5 cover this.

Full topic list: Domain 1 study guide.

What does Domain 2 (Tool Design & MCP Integration) really test?

D2 tests whether you can design tools the model will select correctly and recover from gracefully. Most questions describe a tool set and a symptom (misrouting, retry storms, wrong-customer lookups) and ask which change fixes it.

Representative topics:

• Descriptions decide routing. Vague or overlapping tool descriptions cause misrouting; the fix is targeted rewriting, not more system-prompt keywords. See 30% Misrouting with Minimal Descriptions.
• Error design. Uniform “operation failed” errors waste retries; isError plus a classified error type lets the agent decide whether to retry, escalate or stop. See Structured Errors: 78-95% Recovery vs 15%.
• Tool count and scope. More tools means worse selection; each agent should get only its role’s tools. Also expect MCP scope questions (.mcp.json vs user config, environment-variable expansion for secrets).

Full topic list: Domain 2 study guide.

What does Domain 3 (Claude Code Configuration & Workflows) really test?

D3 is the most “product-specific” domain: the CLAUDE.md hierarchy, rules with path globs, commands vs skills, plan mode, and non-interactive CI usage. Questions tend to be “a team wants X to apply to Y; where does the config live?”

Representative topics:

• Where configuration lives and what actually loads. Put Team Standards Where the Team Can Actually Get Them covers user, project and directory scope; the path-glob articles explain why unscoped rules waste context.
• Commands vs skills vs CLAUDE.md. Commands Are Prompts. Skills Are Configured Workflows. draws the line the exam draws.
• CI and headless runs. The -p flag, structured output flags, and why a fresh session should review code the previous session wrote.

Full topic list: Domain 3 study guide.

What does Domain 4 (Prompt Engineering & Structured Output) really test?

D4 is not “write a nice prompt”. It is about specificity, calibration with examples, guaranteed structure via tool_use, semantic validation, and the Batch API’s cost/latency trade-off.

Representative topics:

• Specific criteria and few-shot calibration. Vague adjectives produce inconsistent output; two or three diverse examples beat paragraphs of instruction. See Text Instructions Failed at 64%. Two Examples Reached 91%.
• Structure vs semantics. tool_use with a schema removes structural errors but not wrong values, so validation and corrective retry still matter. See tool_use Eliminates Structural Errors. Semantic Errors Remain.
• Batch API. Half the cost, up to a day of latency, custom_id for correlation, resubmit only what failed.

Full topic list: Domain 4 study guide.

What does Domain 5 (Context Management & Reliability) really test?

D5 asks what happens to a long-running agent over time: facts get summarized away, attention sags in the middle of the context, tool output bloats, and errors get swallowed. It also covers escalation design and monitoring.

Representative topics:

• Progressive summarization and lost-in-the-middle. See The U-Shaped Attention Curve and the persistent case-facts pattern.
• Escalation triggers. Sentiment and model confidence are unreliable triggers; explicit human requests and policy gaps are the right ones. See Confidence Scores Are Unreliable.
• Structured error context and access-failure vs valid-empty. “No results” after an outage is not the same as “no results exist”.

Full topic list: Domain 5 study guide.

What the exam is really testing across all five. Almost every question is a scenario with constraints and four plausible designs; you pick the least bad one. The recurring theme is knowing which layer a guarantee belongs to: prompts for preferences, hooks and schemas for hard rules, tool design for routing and recovery, context structure for long sessions. If you can explain why a wrong option is wrong in those terms, you are ready for that domain.

Next step

Skim the full study guide, then take the free 60-question mock exam, which samples questions by official domain weight so your first score tells you where you stand per domain. Start deep study with Domain 1; it is the biggest slice and its ideas reappear everywhere else.