クロード認定アーキテクト基礎（CCA-F）模擬試験

Introduction As the business utilization of generative AI rapidly advances, Anthropic's "Claude" is garnering overwhelming support in the enterprise sector. As an official certification to prove the ability to implement its advanced prompt design and agent architecture, the "Claude Certified Architect Foundations (CCA-F)" has emerged. This exam proves that you are a "practical architect" capable of building production-level AI applications making full use of Claude Code, the Agent SDK, and MCP (Model Context Protocol), rather than just having general AI knowledge. Benefits of Starting to Study for CCA-F Now Currently, eligibility for the CCA-F exam is restricted to access from the "Claude Partner Network," meaning only those belonging to Anthropic's partner companies can take it. However, because Anthropic moves very quickly, it is highly anticipated that it will be opened to the general public in the near future. Therefore, even for those not working at a Claude partner company, starting to study now offers the following benefits: A Head Start upon General Release: When the exam is released to the public, you can immediately obtain the certification and establish a competitive edge in the market. Deep Understanding of Claude and Direct Link to Practice: Beyond simply aiming for certification, learning best practices for "Agentic Architecture," "MCP Integration," and "Advanced Prompt Engineering" through your studies will dramatically improve your baseline understanding of Claude and your practical application level. In this article, in addition to an overview of and strategies for the CCA-F exam, we will introduce a Udemy practice exam that provides an "extremely efficient learning route" to passing the test. Overview of the Claude Certified Architect Foundations (CCA-F) Exam For those aiming to get certified, let's first summarize the basic information of the CCA-F exam. Basic Exam Information The specifications for the CCA-F exam are as follows: Exam Duration: 120 minutes Number of Questions: 60 questions (Multiple choice and multiple response) Passing Score: 720 / 1000 (Approx. 72% accuracy) Exam Fee: 99 USD (excluding tax) Language: English only Exam Method: Online (Remote proctored) Certification Validity: 6 months Exam Scope and Weighting This exam consists of the following 5 domains, each with a different weighting. Agentic Architecture & Orchestration: 27% This is the most heavily weighted area. It tests your design capabilities in orchestrating multi-agents using the Agent SDK, hub-and-spoke architecture, and controlling agent loops. Claude Code Configuration & Workflows: 20% It tests your practical configuration skills of Claude Code in development workflows, such as configurations using CLAUDE.md, Agent Skills, and integration into CI/CD pipelines. Prompt Engineering & Structured Output: 20% Tests your advanced prompting techniques for reliable structured data extraction using JSON schemas, few-shot prompting, and obtaining accurate outputs while suppressing hallucinations. Tool Design & MCP Integration: 18% This area relates to building MCP (Model Context Protocol) servers and defining tools. It covers designing tool boundaries, writing accurate descriptions, and resource management. Context Management & Reliability: 15% Questions cover managing long contexts, countermeasures for the "Lost in the middle" problem, appropriate context passing between multi-agents, and error handling practices. Features of This Practice Exam The actual exam is available only in English. Reading and deciphering specialized, lengthy scenario questions related to AI in English is a significant burden for non-native speakers. Therefore, this practice exam adopts a unique structure to help you reliably build your skills while overcoming the language barrier. Detailed Explanations for Each Option In scenario questions, it is important to distinguish between "similar approaches." We explain in detail from an architectural perspective not only the correct answer, but also why the other methods are not optimal for that specific case. Provision of Relevant Official Documentation For complex configurations and advanced specifications, we have comprehensively included links to Anthropic's official documentation so you can check them immediately. Sample Questions We have included examples of actual questions and their explanations for your reference. Question 1 A supply chain management platform uses a "Tracking Agent" that scans thousands of shipping logs and a "Logistics Strategy Agent" that builds timelines of events. The Tracking Agent outputs summaries of relevant logs to the Logistics Strategy Agent. However, the Logistics Strategy Agent frequently gets the chronology of events wrong and cannot show users which logs support specific claims, resulting in an inaccurate timeline. How should the architect fix this multi-agent workflow? Options Instruct the Logistics Strategy Agent to infer the chronological order based on the context in the summary. Increase the Temperature of the Logistics Strategy Agent to improve its ability to creatively synthesize a timeline from unstructured data. Have the Tracking Agent output structured JSON that explicitly includes metadata such as precise timestamps, sender/receiver IDs, and source file names alongside the extracted facts. Combine the Tracking Agent and the Logistics Strategy Agent into a single prompt so the model can read shipping logs and build the timeline simultaneously. Overall Explanation When passing information between agents, relying solely on descriptive natural language summaries (narrative summaries) carries the risk of omitting important metadata such as dates and IDs, or having the order swapped during the model's reasoning process. When high-precision timeline construction and evidence presentation are required, it is essential to explicitly pass metadata using structured data (JSON). This enables downstream agents to perform accurate sorting and referencing of sources (provenance) based on objective numerical data. Explanations for Each Option 1. Incorrect Relying on inference makes it prone to hallucinations, and especially when handling large amounts of data, it is not a fundamental solution to prevent sequencing errors. 2. Incorrect Increasing the temperature increases the randomness of the output, which is counterproductive because it compromises the "accuracy" that is most important in creating a fact-based timeline. 3. Correct By passing metadata as structured data, the Logistics Strategy Agent can accurately organize the chronological order and specify sources based on reliable data without relying on guesswork. 4. Incorrect Attempting to process thousands of shipping logs at once will cause context window limitations and information loss, paradoxically degrading accuracy, making it an inappropriate architecture. 【official documentation】 Building effective agents Question 2 A financial analyst is building a bot to automate investment research. When the agent calls a stock data retrieval tool (when stop_reason is tool_use), the API response includes Claude's reasoning process (a text block) followed by the tool execution instruction (a tool_use block). To save token usage and conserve context, the developer deleted this reasoning text block and saved only the tool_use block in the conversation history. What is the impact of this optimization on the architecture? Options Claude will automatically regenerate the missing text block in the next turn, so latency will increase but reasoning will not be affected. The API will reject subsequent requests because there is a restriction that a text block must always precede a message containing a tool_use block. Claude will lose the context of its Chain of Thought, degrading its ability to judge why it called the tool and how it should interpret the subsequent results. This is a recommended optimization technique that can reduce token usage without affecting the model's reasoning capabilities. Overall Explanation The text block output by Claude when using a tool functions as a "Chain of Thought" for the model to organize its thinking process. If this reasoning portion is deleted from the conversation history, the model loses the logical context of "why it called that tool" and "how it should interpret the obtained results." As a result, its reasoning ability degrades in subsequent turns, increasing the risk of drawing inaccurate conclusions. Maintaining the model's logical consistency is more important than saving tokens. Explanations for Each Option 1. Incorrect Claude only references the history provided as a prompt. It does not have a feature to automatically supplement or regenerate content deleted in past turns, and the context remains lost. 2. Incorrect Technically, the API will not reject the message (it is allowed as a message list structure), but practical problems will occur because the model's logical consistency is compromised. 3. Correct The reasoning process contained in the text block is crucial contextual information for the model to correctly execute complex tasks, and deleting this leads to intentionally limiting the model's "intelligence." 4. Incorrect No, this is an anti-pattern. It is an act of sacrificing the model's reasoning accuracy for trivial token savings, and it should be avoided as it significantly impairs system reliability. 【official documentation】 Tool use with Claude Question 3 A market research agent is trying to extract (scrape) pricing information from a specific competitor's website. The site's firewall detected the agent and blocked its IP address. How should the sub-agent construct the error response to the coordinator? Options Return a structured error detailing the IP block, the attempted URL, and suggesting delegating the task to a sub-agent with residential proxy capabilities. Return an empty list indicating that information could not be extracted from that site. Return a generi