Claude API Key to CCA-F: What the Exam Actually Tests

If you have ever typed your Claude API key into a project for the first time, you already know the gap between "it works" and "it works reliably at scale." The Claude Certified Architect, Foundations exam (CCA-F) is Anthropic's attempt to measure that gap. Launched on 12 March 2026, it is the first professional certification in the Claude Partner Network, a $100M programme that had already attracted more than 40,000 partner applicant firms and 10,000 certified individuals as of 3 June 2026.

This guide answers the questions that come up most often from solution architects comparing the CCA-F to existing cloud certifications, deciding where to focus their study time, and wondering whether the credential is a genuine architecture test or a vendor-specific product quiz.

How does the CCA-F compare to AWS, Azure, and GCP architect certifications?

The CCA-F is narrower in infrastructure scope than any of the big-three cloud architect exams, but it goes considerably deeper on AI system design. AWS Solutions Architect Professional, for example, spans networking, storage, databases, and compute across hundreds of services. The CCA-F covers exactly one model family across five domains, but the questions are almost entirely scenario-based judgment calls rather than service-feature recall.

The more useful comparison is to AWS Machine Learning Specialty or Azure AI Engineer Associate. The CCA-F is closer to those in spirit, but it is more architecture-oriented than either: it expects you to design multi-agent systems, reason about failure modes, and choose between orchestration patterns, not just call a managed API.

Is the exam really scenario-heavy, and how much judgment does it test?

Every one of the 60 questions is scenario-based multiple choice, with one correct answer and three plausible distractors. "Plausible distractors" is the operative phrase. Anthropic has designed the exam to reward three consistent principles: prefer deterministic solutions over probabilistic ones when stakes are high, apply proportionate fixes rather than over-engineering, and trace to root causes rather than treating symptoms.

That means a question about a flaky agentic loop will not ask you to name a class or recite a parameter. It will describe a production scenario, give you four architecturally coherent options, and expect you to identify which one addresses the actual failure mode. Concepts like agentic loop anti-patterns and debugging premature loop termination appear precisely because the exam tests whether you can diagnose before you prescribe.

The scored scale runs from 100 to 1000, with 720 as the passing score. Anthropic does not publish the raw-to-scaled conversion, so we will not speculate on an exact question count for passing.

Which domains carry the most weight, and where should architects focus?

The five domains and their exam weights are:

Domain 1 is the largest single block. If you are a solution architect who has spent time designing distributed systems, the concepts will feel familiar in structure but unfamiliar in specifics. The exam tests things like hub-and-spoke architecture for multi-agent systems, coordinator responsibilities, parallel subagent spawning, and how to handle structured context passing between agents without losing attribution.

Domains 3 and 4 are tied at 20% each, which surprises many candidates who expect prompt engineering to be a minor topic. In practice, the exam treats structured output design and prompt construction as architectural decisions, not afterthoughts. A poorly designed prompt schema is a reliability risk, and the exam tests it accordingly.

Domain 5 at 15% covers context management, session handling, and reliability patterns. It is the smallest domain by weight but arguably the most operationally consequential: context failures in production are silent and expensive.

What does the exam test in Agentic Architecture specifically?

Domain 1 covers 27% of the exam across a range of task statements that go well beyond "understand what an agent is." The exam expects you to reason about:

1. When to use model-driven versus pre-configured decision making in an orchestration layer.
2. How to design prerequisite gate design so that downstream agents do not execute on incomplete inputs.
3. How subagent context isolation prevents cross-contamination in parallel workloads.
4. When to use fork_session for divergent exploration versus resuming an existing session.
5. How to implement multi-agent error handling and routing without masking the root cause.

The exam consistently rewards candidates who can identify the narrowest intervention that fixes a problem. A question about an agent that silently drops results will not reward "add more logging everywhere." It will reward identifying the specific point in the pipeline where attribution is lost, which is the kind of reasoning covered in diagnosing attribution loss in synthesis.

How deeply does the exam test Claude Code and CLAUDE.md workflows?

Domain 3 at 20% is more operationally specific than most candidates expect. It covers the three-level configuration hierarchy for Claude Code, CLAUDE.md structure, hooks, skills, and the version control implications of different configuration choices.

This is not a general LLM design domain. It tests whether you understand how Claude Code reads configuration at the project, user, and enterprise levels, how hooks intercept tool calls before and after execution, and how skills differ from commands in their scope and persistence. Candidates who have only used Claude via the API and have never configured a Claude Code workspace will find Domain 3 the steepest learning curve.

The practical implication: if your Claude API key usage has been entirely programmatic, budget extra study time for Domain 3 before the exam.

How important is prompt engineering versus architecture for passing?

Domain 4 at 20% means prompt engineering is not a minor topic, but the exam does not test it in isolation. It tests prompt engineering as a component of system design. Questions in this domain typically present a pipeline that is producing inconsistent or malformed output and ask you to identify the highest-leverage fix.

The exam's preference for proportionate fixes is particularly visible here. A question about inconsistent JSON output will not reward "switch to a different model." It will reward identifying whether the schema definition, the few-shot examples, or the output validation step is the weak link, and fixing that specific link.

Structured output design, schema construction, and the interaction between system prompts and tool descriptions are all testable. The tool descriptions as selection mechanism concept, for example, sits at the intersection of Domain 2 and Domain 4: a poorly written tool description is both a routing failure and a prompt engineering failure.

Is context management a major exam theme or a secondary topic?

At 15%, Domain 5 is the smallest domain, but "secondary" would be the wrong word. Context management failures are among the most common production failure modes in Claude deployments, and the exam reflects that. The domain covers stale context problems, summary injection for fresh sessions, and the decision logic for when to resume versus fork versus fresh start.

RAG and caching are relevant insofar as they affect context quality and token efficiency, but the exam does not test them as standalone infrastructure topics. The focus is on how context degrades over extended sessions, how to detect that degradation, and how to recover from it without discarding useful state.

Candidates who have worked with long-running agentic workflows will recognise the failure modes immediately. Those who have only built stateless API integrations should treat Domain 5 as a domain that requires deliberate study, not a domain to skim because of its lower weight.

Is the CCA-F a hands-on architect credential or a vendor skills check?

The honest answer is that it is both, and the distinction matters less than it might seem. The exam is vendor-specific by definition: it tests Claude APIs, MCP, and Claude Code, not LLMs in general. But the reasoning it rewards is genuinely architectural. Candidates who pass by memorising Claude-specific parameters without understanding why those parameters exist will struggle with the scenario questions.

The more useful framing is this: the CCA-F is a vendor-specific exam that tests architecture-quality reasoning within that vendor's ecosystem. It is closer in spirit to an AWS Solutions Architect exam than to an AWS product certification like "AWS Certified Alexa Skill Builder." The 30 task statements across five domains are designed to test whether you can build reliable systems with Claude, not whether you can recite the API surface.

As of 3 June 2026, more than 10,000 individuals have passed the exam. Anthropic has announced further architect, developer, and seller certifications planned for later in 2026, which suggests the CCA-F is the foundation of a broader credential stack rather than a standalone product quiz.

How does practice exam preparation compare to the real exam?

The CCA-F uses the same 100-to-1000 scale and 720 passing bar as the real exam, which means a well-constructed practice exam can give you a calibrated signal on readiness. The key variable is question quality: practice questions that test recall will not prepare you for scenario questions that test judgment.

At AI Skill Certs, our practice exams are 60 questions scored on the same scale, and our adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold before it considers a concept covered. That means the engine will not move you past tool design and MCP integration until your response pattern demonstrates genuine understanding, not lucky guessing. We are an independent platform, not affiliated with or endorsed by Anthropic.

The concepts library at /concepts covers 174 atomic concepts mapped to all five domains and all 30 task statements. For candidates who want to understand why an answer is correct rather than just which answer is correct, that mapping is the most direct path to exam-quality reasoning.

What is the most efficient study sequence for a solution architect?

Given the domain weights and the scenario-heavy question style, we recommend the following sequence:

1. Domain 1 first (27%): Spend the most time here. Work through agentic architecture patterns, coordinator design, and multi-agent error handling before touching other domains.
2. Domains 3 and 4 together (20% each): Claude Code configuration and prompt engineering reinforce each other. Study them in parallel rather than sequentially.
3. Domain 2 (18%): Tool design and MCP integration build on Domain 1 concepts. Study it after you have a solid foundation in orchestration.
4. Domain 5 last (15%): Context management concepts are easier to absorb once you understand the session lifecycle from Domains 1 and 3.

The exam costs $99 per attempt. Tiered Anthropic partners receive a discounted first attempt. Given the scenario-based format, a second attempt is not a reliable fallback strategy: the same judgment gaps that caused a first failure will recur unless you address the underlying conceptual weaknesses.