The Ultimate Study Guide for the Claude Certified Architect (CCA-F) Exam

Mastering the Claude Certified Architect (CCA-F) Exam

Preparing for the Claude Certified Architect (CCA-F) exam requires a deep understanding of agentic architecture, tool design, and context management. The 2026 exam consists of multiple-choice questions focused on high-signal, actionable intelligence. To pass, you need to achieve a target score of 720/1000.

This complete visual playbook breaks down the critical domains, orchestration loops, and anti-patterns you must know to succeed.

The Architect's Mindset: 3 Core Philosophies

Before diving into technical configurations, it is crucial to adopt the three core philosophies that govern Claude architecture:

*
Guarantees Beat Instructions: When a rule must absolutely hold, enforce it in code using hooks, gates, or schemas. Never rely solely on a prompt for high-stakes security boundaries, as prompts are probabilistic while code is deterministic.

*
Fix the Root Cause Proportionately: The exam will constantly offer over-engineered distractors. If two tools are getting confused, improve their descriptions first before attempting to build a complex routing classifier.

*
Context is Scarce and Leaky: The working memory of the model fills quickly, and middle context often gets lost. Always extract key facts and pass vital information explicitly.

Exam Blueprint & Domain Breakdown

The CCA-F exam is distributed across five main domains:

*
Domain 1: Agentic Architecture & Orchestration (27%)

*
Domain 3: Claude Code Configuration (20%)

*
Domain 4: Prompting & Structured Output (20%)

*
Domain 2: Tool Design & MCP (18%)

*
Domain 5: Context Management (15%)

Because orchestration and tool design make up 45% of the exam, you should master these core technologies first. Agents are only as reliable as the structure built around them.

Orchestration & Execution Loops

Understanding how agents loop and use tools is a major focus. During the execution loop, if the `stop_reason` is `tool_use`, the system must run the tool, append the result, and loop back. Claude has no outside memory, so results must be appended. The loop only exits when the `stop_reason` is `end_turn`. Never rely on text-parsing or an iteration cap to end the loop.

In a Hub & Spoke architecture, a Coordinator decomposes tasks, delegates to specialists, aggregates results, and evaluates for gaps. Specialist sub-agents have no shared memory, meaning context must be passed explicitly. If sub-agents succeed but the final output has massive holes, the Coordinator likely sliced the task too thin.

Tool Design & Enforcement Hooks

The exam distinguishes heavily between probabilistic LLM mechanisms (like prompt engineering) and deterministic code mechanisms. If a question demands 100% compliance, the answer is always a deterministic Hook.

*
Pre-Tool Use Hook: Intercepts and blocks actions before the tool runs, such as stopping any refund over a certain limit regardless of the prompt.

*
Post-Tool Use Hook: Normalizes messy output after the tool returns data.

Error Recovery Matrix

When handling tool design, you must understand how to architect recovery based on the error type:

*
Transient Errors (e.g., Timeout): Require an automatic retry.

*
Validation Errors (e.g., Malformed ID): The agent must fix the input format.

*
Business Errors (e.g., Refund exceeds policy): Unretryable; the system should explain the policy or offer an alternative.

*
Permission Errors: Escalate to a human.

Context Leakage & Escalation Protocols

To prevent facts from being lost through progressive summarization, move hard facts like IDs and amounts out of the summarized history and into a structured "Case Facts Block" that rides alongside every prompt.

When deciding to escalate to a human, always honor explicit customer requests, escalate policy gaps, and escalate permission blocks. Never hand a human a "cold start". Always pass a structured summary containing the Customer ID, recommended action, and prior steps.

Identifying Exam Distractors (Anti-Patterns)

You can instantly spot incorrect multiple-choice answers by recognizing these common anti-patterns:

*
Over-Engineering: Do not select answers that suggest building an ML routing classifier when tools get confused; instead, fix the root cause proportionately by improving tool descriptions.

*
More Prompting: Rewording a prompt to enforce a strict financial limit will fail; it requires a deterministic hook.

*
Invented Features: Eliminate options that suggest made-up CLI flags.

By mastering these mechanics and focusing your first five days of study on Orchestration and Tools, you will be well-prepared to pass the CCA-F exam.