CCA-F Exam Study Guide: Claude Certified Architect Foundations

June 23, 2026 · updated June 25, 2026 · How Claude Actually Works (part 32)

▶ Watch on YouTube & subscribe to The Stack Underflow

Anthropic launched the Claude Certified Architect: Foundations exam — abbreviated CCA-F — in March 2026. It is a 60-question, scenario-driven assessment of whether you can make sound architectural decisions when building production systems on the Claude API, Claude Code, the Agent SDK, and MCP. The “F” stands for Foundations, but do not let that word mislead you: the exam tests applied judgment in realistic production trade-off scenarios, not definition recall.

The most useful insight about this certification is structural: its five exam domains are a direct projection of the Claude stack you have been studying throughout this series. Each domain corresponds to a specific layer or cross-cutting plane. Once you see the mapping, you stop treating the exam as an opaque question bank and start treating it as a weighted audit of your stack knowledge — which is exactly what efficient preparation requires.

The one-sentence version: Every CCA-F domain maps to a named layer or plane of the Claude stack, so weighting your study hours by domain score percentage is the fastest path to passing.

The Claude Stack as an Exam Blueprint

Before the domain mapping makes sense, you need the five-layer mental model this series calls “the Claude stack.” Two additional cross-cutting planes — prompts and reliability — run horizontally across every layer rather than belonging to any single one.

┌───────────────────────────────────────────────┐
│  L4  Surfaces     Claude Code, Claude.ai, etc │
├───────────────────────────────────────────────┤
│  L3  Orchestration  Agent SDK / multi-agent   │
├───────────────────────────────────────────────┤
│  L2  Reach          Tools + MCP servers       │
├───────────────────────────────────────────────┤
│  L1  Protocol       Messages API              │
└───────────────────────────────────────────────┘
  ←────── Prompts plane (cuts across all L) ──────→
  ←── Reliability plane (cuts across all L) ──────→

The exam has exactly five domains, and each one lands on a clearly identifiable part of this diagram. That is not a coincidence — the certification was designed to audit competency across the complete stack.

Domain-to-Stack Mapping

The table below is the core reference for your prep. The “Stack Home” column tells you which part of the diagram a domain lives in; the “What it tests” column tells you the architectural judgment calls you need to make under exam conditions.

The two planes — D4 (prompts) and D5 (reliability) — appear as independent domains precisely because they are not tied to any single layer. Context management matters at L1 (token budget in the API call), at L2 (MCP tool responses consuming tokens), at L3 (multi-turn agent history), and at L4 (Claude Code’s conversation compaction). Treating D5 as a narrow “ops checklist” is a common prep mistake.

Score Weights: Where to Spend Your Time

The domain weights are the most actionable data point in the exam blueprint. Internalize this distribution before you open a single study resource:

D1  Agentic & Orchestration   ████████████████████████  27%
D3  Claude Code               ███████████████           20%
D4  Prompt Engineering        ███████████████           20%
D2  Tool Design + MCP         █████████████             18%
D5  Context + Reliability     ███████████               15%
                                                       ────
                                                        100%

D1 + D3 + D4 = 67% of the total score.

Two-thirds of your points live in three domains. That is the lever. If you have limited study time, depth in D1, D3, and D4 returns far more points per hour than breadth across all five.

D2 at 18% is still nearly one-fifth of the exam — and it underpins D1. Agent orchestration scenarios routinely require reasoning about how tools and MCP servers connect. Skipping D2 creates cascading gaps in D1 answers.

Exam Mechanics

The mechanics are fixed and worth committing to memory, because misunderstanding them leads to poor test-day pacing.

The most consequential of these is one attempt. There is no “test the waters” run. Build a readiness signal before you register: use Anthropic’s official practice exam, not third-party question banks, which are less calibrated to the scenario format.

The exam rewards architectural judgment — the ability to look at a realistic production situation, evaluate four plausible options, and identify the one that is correct under the stated constraints. Two of the four options will usually look reasonable. You need to reason about trade-offs, not recall a flag name.

Your Study Path by Domain

The video maps each domain to specific episode blocks in this series. Use this table as your study scheduler:

Study in weight order, not domain-number order: D1 first, then D4, then D3, then D2, then D5. The capstone exercises (07-02 through 07-04) are scenario-based simulations of the exact judgment-call format the exam uses — work through all three before you sit.

A Note on Fast-Moving Facts

Several things the exam can touch are actively evolving as of June 2026:

Current flagship models (docs.anthropic.com, June 2026):
  claude-opus-4-8     — highest capability, 1M context window (GA)
  claude-sonnet-4-6   — balanced, 1M context window (GA)
  claude-haiku-4-5    — fastest / lowest cost

Context windows: 1M tokens at standard pricing for Opus 4.x / Sonnet 4.x
  (long-context pricing premium removed in March 2026)

Prompt caching: cache_control breakpoints, up to 90% cost reduction
  on cached prefix tokens; combinable with Batch API for ~95% reduction

stop_reason values (Messages API):
  end_turn | max_tokens | stop_sequence | tool_use | refusal
  (stop_details object added in Opus 4.7 for refusal classification)

The exam blueprint was written against the architecture as of its launch. Verify current model IDs, pricing, and context-window specs against docs.anthropic.com before you sit — Anthropic updates these regularly, and exam questions are revised to stay current.

Common Misconceptions

“I can cram config details the night before.” The exam is explicitly scenario-based. The question is never “what is the name of this flag?” — it is “given these production constraints, which of these four architectural decisions is correct and why?” Configuration memorization without architectural understanding fails in the two-option-looks-right scenarios that dominate the exam.

“Domain 2 (Tool Design and MCP) matters less because it has the lowest weight.” D2 at 18% is the fourth-ranked domain, not the fifth — D5 (15%) is smaller. More importantly, tool design and MCP integration are foundational to D1 (agentic orchestration). Agent scenarios require you to reason about how tools are handed to subagents, how MCP servers distribute capabilities, and what happens when a tool call fails. Weak D2 preparation creates visible gaps in D1 answers.

“The reliability domain is just about retries.” D5 covers context-window budget management, prompt-caching strategy (cache_control breakpoints, cache invalidation), conversation compaction patterns, evaluation design, and safe degradation paths when an agent hits a capability boundary. It is a systems-thinking domain — “retries” is one bullet point among many.

“If access is partner-restricted, I can’t start studying now.” Study material is fully available regardless of your registration status. The 05-series (Claude Code), 04-series (prompts), and 03-series (Agent SDK) are accessible now. Build competency first; sort registration logistics when Anthropic opens public access. Waiting on access to start studying is leaving free runway on the table.

Frequently Asked Questions

Who is the CCA-F designed for? Developers and architects who build production systems on the Claude API, Claude Code, the Agent SDK, and MCP. It is a foundations-level certification, so it assumes you have working familiarity with the API and can already write agentic code — it does not assume research-track or ML expertise. The exam tests whether you can design and operate these systems well at scale, not whether you can implement model training.

Is the exam open to the public? As of June 2026, the CCA-F is accessible through the Claude Partner Network. The first 5,000 Partner Network seats were offered free; everyone else pays $99 per attempt. Anthropic has announced public registration is coming but has not confirmed a date. Verify current access requirements directly on Anthropic’s official site before planning your attempt — the program is actively evolving.

What is the best readiness signal before I register? Anthropic’s official practice exam is the most accurate signal available. Third-party question banks exist but are less calibrated to the scenario-based format the actual exam uses. A good self-check heuristic: if you can read through capstones 07-02, 07-03, and 07-04 and explain why the correct answer is correct for each scenario, you are ready. If you are guessing or uncertain on more than a quarter of those, keep studying.

Should I study domains in order from D1 to D5? Study by weight, not by number. The recommended ordering for maximum ROI is D1 (27%) first, then D4 (20%), then D3 (20%), then D2 (18%), then D5 (15%). If your available time is constrained, stopping after D3 still gives you coverage of 67% of the total score. Add D2 and D5 as time permits.

What does “architectural judgment” actually mean in practice? The exam presents a realistic production situation — for example, a multi-agent pipeline where a subagent needs to call an external API, the orchestrator needs to handle failures gracefully, and context budget is a constraint. You are given four design options. Two look plausible. You must identify which one correctly balances the given constraints. Getting this right requires understanding why design patterns exist, not just what they are. That is what makes the capstone exercises (07-02 through 07-04) the most direct prep: they drill exactly this reasoning format.

How long does the exam take to prepare for? This varies by starting point. If you have worked through the full series and completed the capstone exercises, you are close. Plan for a focused 2–3 week review sprint: map your weak domains using the domain table above, prioritize the corresponding episode blocks, then use the official practice exam to gauge readiness. Do not book the exam before you have a passing score on the practice run — one attempt is all you get per registration.

Where This Fits in the Series

This tutorial is the series terminus for How Claude Actually Works. It does not introduce new technical concepts — instead, it shows how the concepts from every prior episode map onto a formal certification framework. Think of it as the syllabus review at the end of the course.

The underlying architecture you need to understand for this exam is covered across the series. The Claude stack mental model is the foundation every domain builds on. The MCP and tooling domain (D2) is covered in what is MCP and how Claude uses tools. Claude Code (D3) runs deep across how Claude Code works, Claude Code hooks explained, and Claude Code skills, subagents, and hooks. The prompts and reliability domains (D4 and D5) draw on prompt caching, context engineering, and how to write LLM evals.

Browse all tutorials to follow the full series from the beginning.