CCAR-F study guide: domains, scenarios and a 4-week plan

Claude Certified Architect: Foundations (CCAR-F) is Anthropic's certification for solution architects who make design decisions when building production applications with Claude Code, the Claude Agent SDK, the Claude API and the Model Context Protocol (MCP). Every question sits inside a realistic scenario and asks for the best trade-off, not a definition. You answer 60 multiple-choice and multiple-response questions in 120 minutes, built around 4 scenarios drawn at random from a published set of 6.

CCAR-F at a glance

Details as of September 2026 — confirm on the official exam page before booking.

The credential is valid for 12 months, with a free on-time renewal assessment on the Anthropic Partner Academy.

Who this exam is for

The exam guide describes a solution architect with hands-on experience of multi-agent orchestration and hooks in the Agent SDK, configuring Claude Code for a team, designing MCP tools and resources, producing reliable structured output, managing context, and wiring Claude into CI/CD. It suggests six months or more of practical building with these tools. There is no formal prerequisite.

Anthropic's certification FAQ states that exams are open to people at Claude Partner Network organisations, registering with a partner email address on a recognised company domain.

How CCAR-F differs from the other Claude certifications. The four exams are separate, and none requires another.

• [Associate: Foundations (CCAO-F)](/blog/ccao-f-study-guide) is for professionals using Claude's apps, with no coding.
• [Developer: Foundations (CCDV-F)](/blog/ccdv-f-study-guide) covers the same technologies from the builder's side, with a third of its weight on API mechanics and general software engineering.
• Architect: Foundations (this exam) asks which design choice is right inside a concrete production scenario. It goes deeper than CCDV-F on agent orchestration, Claude Code configuration and structured extraction, and it publishes an explicit out-of-scope list.
• [Architect: Professional (CCAR-P)](/blog/ccar-p-study-guide) widens to enterprise architecture: retrieval pipelines, governance and compliance, evaluation programmes and stakeholder management.

What the exam covers

Knowing the scenarios in advance is the biggest advantage the guide gives you. Each frames a block of questions, and each exam uses four:

1. Customer support resolution agent built on the Agent SDK with custom MCP tools for customer lookup, orders, refunds and escalation.
2. Code generation with Claude Code: slash commands, CLAUDE.md, and plan mode versus direct execution.
3. Multi-agent research system: a coordinator delegating to search, analysis, synthesis and reporting subagents.
4. Developer productivity tools using the built-in tools (Read, Write, Bash, Grep, Glob) and MCP servers.
5. Claude Code in CI/CD: automated review, test generation and pull request feedback with few false positives.
6. Structured data extraction validated against JSON schemas and fed to downstream systems.

The five domains below are what the questions inside those scenarios measure.

Agentic Architecture & Orchestration (27%)

The heaviest domain. Know the agentic loop precisely: continue while the stop reason is tool use, stop at end of turn, and append tool results to history. Know that subagents start with isolated context, so the coordinator must pass everything they need. The guide leans on enforcement: when a rule must hold, such as verifying identity before a refund, use a hook or a programmatic gate rather than a prompt. Task decomposition and session resumption also appear.

Tool Design & MCP Integration (18%)

Tool descriptions are how the model chooses tools, so vague or overlapping descriptions cause misrouting. Expect questions on structured error responses (telling a transient failure from a validation or permission error, and a failed lookup from a genuinely empty result), on giving each agent only the tools its role needs, on tool_choice settings, on MCP server scoping in Claude Code, and on picking the right built-in tool.

Claude Code Configuration & Workflows (20%)

The CLAUDE.md hierarchy and what each level shares, modular rules in .claude/rules/ with path scoping, commands and skills (including running a skill in a forked context and restricting its tools), plan mode versus direct execution, iterative refinement with examples and tests, and running Claude Code non-interactively in CI with structured JSON output.

Prompt Engineering & Structured Output (20%)

Explicit criteria beat vague instructions such as "be conservative". Few-shot examples are the guide's preferred tool for consistency and ambiguous cases. Structured output through tool use and JSON schemas removes syntax errors but not semantic ones, so you still validate and retry with specific feedback. It also covers batch processing trade-offs and why an independent reviewer instance catches more than self-review.

Context Management & Reliability (15%)

Preserving key facts across long conversations (and the risks of progressive summarisation), escalation and ambiguity handling, error propagation between agents, managing context when exploring large codebases, human review with calibrated confidence, and keeping source attribution intact when combining findings from several sources.

A 4-week study plan

Orchestration gets a week of its own; the other domains are paired by the scenarios they share.

Week 1: Orchestration (27%). Build a coordinator with two subagents in the Agent SDK. Pass context explicitly, run subagents in parallel, and add a hook that blocks a tool call until a prerequisite has succeeded.

Week 2: Tools and MCP (18%) with context and reliability (15%). Write three tools with deliberately overlapping descriptions, watch the misrouting, then fix the descriptions. Return structured errors from an MCP tool and have the agent respond differently to each type. Practise keeping transactional facts intact across a long conversation.

Week 3: Claude Code (20%). Configure a repository with user, project and directory CLAUDE.md files, a path-scoped rule and a skill. Use plan mode on a multi-file change and direct execution on a one-line fix. Run Claude Code in a CI script with JSON output against a schema.

Week 4: Prompting and structured output (20%), then rehearsal. Build an extraction tool with a schema that includes nullable fields and an "other" option, add a validation-and-retry loop, and write review criteria that reduce false positives. Then sit a timed mock and study your weakest scenario.

Where candidates lose marks

• Prompts instead of guarantees. When the scenario says a rule must always hold, answers that add stronger wording to the system prompt are distractors. The guide favours hooks and programmatic gates for deterministic compliance.
• Assuming subagents share memory. A subagent does not see the coordinator's conversation. Options that rely on implied context are wrong; the fix passes findings explicitly, ideally with source metadata.
• Crude loop termination. Stopping when the model's text "sounds finished" or relying on an iteration cap as the main exit are named anti-patterns. The stop reason decides.
• Too many tools per agent. Giving every agent every tool degrades selection. Scoped tool sets, with narrowly shared cross-role tools where needed, are the pattern the guide rewards.
• Studying out-of-scope material. The guide lists exclusions, including fine-tuning, computer use, vision, streaming implementation, rate limits and pricing, cloud provider configuration, and prompt caching beyond knowing it exists. Time spent there does not score.
• Guide terms that have moved. The guide calls the subagent-spawning tool Task; current Agent SDK documentation names it Agent. Claude Code's MCP configuration also has a local scope, the default, alongside project and user. Know both vocabularies.

Practising for CCAR-F on CertifyAIx

CertifyAIx offers original scenario-style questions written against Anthropic's exam guide, covering every domain. The [CCAR-F practice questions](/certifications/claude-certified-architect-foundations) page includes:

• 1,982 original questions across all five domains and every task statement
• 3 full-length mock exams with 60 questions in 120 minutes, the same as the real exam
• 82 mixed domain papers and 90 topic papers of up to 25 questions each (60 minutes per 25)
• An explanation for every question, covering why each option is right or wrong

See [how CertifyAIx practice questions are written](/blog/how-certifyaix-practice-questions-are-written) and [how to use practice exams effectively](/blog/how-to-use-practice-exams-effectively).

Frequently asked questions

How hard is CCAR-F?

It rewards production judgement more than recall. Candidates who know the tools only from documentation tend to struggle where every option is technically possible and one is best.

How long should I prepare?

Four to eight weeks is a reasonable range. The lower end suits someone already running Agent SDK and Claude Code work in production; the upper end leaves room to build the scenarios yourself if parts of the stack are new.

Are CertifyAIx questions real exam questions?

No. Every CertifyAIx question is original. Exam content is confidential under Anthropic's non-disclosure agreement, and misconduct can lead to a revoked credential.

Will I see all six scenarios?

No. Each exam presents 4 scenarios chosen at random from the 6 in the guide, so prepare for all of them.

What is the retake policy?

You wait 14 days after a first failure, 30 after a second and 90 after a third, with up to four attempts in a rolling 12 months. Each attempt costs the full fee.

CertifyAIx is independent and not affiliated with Anthropic. Claude Certified Architect: Foundations is a trademark of its owner. All CertifyAIx practice questions are original.