Anthropic Claude Certification Exam: Complete 2026 Overview

The anthropic claude certification exam known as the Claude Certified Architect, Foundations (CCA-F) launched on 12 March 2026 as Anthropic's first professional certification. It is a 60-question, scenario-based assessment scored on a 100-to-1000 scale, with 720 as the passing mark. This post maps every structural detail you need before you sit the exam, from domain weights to the trade-off logic the questions reward.

What is the format and passing score of the CCA-F exam?

The exam delivers 60 scenario-based multiple-choice questions, each with one correct answer and three plausible distractors. The scoring scale runs from 100 to 1000, and the passing score is 720. Per Anthropic's exam guide, the raw-to-scaled conversion is not published, so we do not state an exact question count as the pass mark; a linear reading puts it at roughly 41 to 42 correct answers, but treat that only as orientation.

The exam is closed-book and available either online-proctored or at a test centre. A single attempt costs $99; tiered Anthropic partners receive a discounted first attempt. Anthropic has not published a formal time limit in the exam guide, but candidates consistently report a 120-minute window.

As of 3 June 2026, more than 10,000 individuals have already earned the certification, part of the broader Claude Partner Network programme backed by $100M in investment.

How are the five domains weighted?

The exam blueprint divides 60 questions across five domains and 30 task statements. Understanding the weights tells you where to concentrate study time.

Domain 1 carries the heaviest weight at 27%, which means roughly 16 of the 60 questions touch agentic architecture topics: agentic loops, subagent spawning, session state, coordinator responsibilities, and multi-agent error handling. Domains 3 and 4 are tied at 20% each, making Claude Code configuration and prompt engineering jointly the second-largest study surface.

What scenario types appear in Domain 1: Agentic Architecture?

Domain 1 is the backbone of the exam. Questions here test whether you can reason about agentic loops, subagent spawning, and session state under realistic production conditions, not just recite definitions.

Agentic loops and stop-reason inspection. A common scenario presents a loop that terminates prematurely. The correct fix traces to inspecting the stop_reason field rather than patching the prompt. Our concept on stop_reason field inspection covers the exact decision logic the exam rewards.

Subagent spawning and context isolation. Questions ask you to choose between spawning parallel subagents and running sequential tasks. The exam rewards parallel subagent spawning when tasks are independent and latency matters, and sequential execution when ordering constraints exist. Subagent context isolation is tested separately: each subagent should receive only the context it needs, not the full coordinator history.

Session management trade-offs. Scenarios present a long-running agent whose context has grown stale. The exam distinguishes between resuming a session, forking it, and starting fresh. The when to resume vs fork vs fresh start concept maps directly to these questions.

Hub-and-spoke vs flat architectures. Multi-agent system design questions test whether you can identify when a hub-and-spoke architecture is appropriate versus a flat peer arrangement. The coordinator's responsibilities, including dynamic subagent selection, are tested explicitly.

What does Domain 2 test about Tool Design and MCP Integration?

Domain 2 (18%) focuses on the practical mechanics of tool design and the Model Context Protocol. The exam consistently rewards structured, deterministic error handling over vague fallback behaviour.

Structured error responses. When a tool call fails, the exam expects you to return a structured error using the MCP isError flag pattern rather than silently suppressing the failure or returning an empty result. The distinction between an access failure and a valid empty result is a recurring test point.

Tool distribution across agents. Questions ask how to distribute tools across a multi-agent system. The exam rewards scoped, role-specific tool sets over giving every agent access to every tool. Tool distribution strategy design and the tool overload problem are both in scope.

Tool descriptions as a selection mechanism. Claude routes to tools based on their descriptions. Scenarios present misrouting bugs and ask for the lowest-effort, highest-leverage fix. Almost always, the answer is rewriting the tool description rather than restructuring the system prompt.

MCP scoping and environment variables. The MCP scoping hierarchy and environment variable expansion in MCP config appear in configuration-focused questions, particularly those involving multi-environment deployments.

How do Domains 3 and 4 test Claude Code and Prompt Engineering?

Domain 3: Claude Code Configuration and Workflows (20%). This domain tests the three-level configuration hierarchy (user, project, and system), version control implications of CLAUDE.md files, and the design of CI/CD-integrated workflows. Questions reward candidates who understand that configuration choices have downstream consequences for reproducibility and team collaboration.

Domain 4: Prompt Engineering and Structured Output (20%). The exam tests structured output schema design, few-shot example construction, and the conditions under which prompt-based enforcement is sufficient versus when programmatic enforcement is required. The high-stakes enforcement decision rule is directly relevant: when stakes are high, the exam rewards deterministic, programmatic solutions over probabilistic prompt-only approaches.

A recurring question type presents a structured output pipeline that occasionally produces malformed JSON and asks for the most reliable fix. The correct answer typically involves schema validation at the output layer, not simply adding more instructions to the prompt.

What does Domain 5 test about Context Management and Reliability?

Domain 5 (15%) covers the mechanics of keeping long-running agents reliable as context grows. The exam tests summary injection for fresh sessions, the stale context problem, and the attention dilution effect that degrades quality when context windows become overloaded.

Scenarios in this domain often present an agent that produces lower-quality outputs after many turns. The exam rewards root-cause tracing: identifying whether the degradation stems from stale facts in the context, attention dilution from an oversized window, or a missing summary injection step. Proportionate fixes are rewarded; wholesale context resets are penalised when a targeted summary injection would suffice.

How do the real-world scenario types map across domains?

The exam uses six recurring scenario archetypes. Each maps primarily to one or two domains but often requires cross-domain reasoning.

The code review scenario deserves special attention. The exam tests whether candidates understand why an independent review session produces higher-quality results than self-review within the same context. A single agent reviewing its own output suffers from confirmation bias baked into the shared context. The correct architecture spawns a separate review instance with a clean context window.

What practical experience does the exam reward?

The CCA-F is positioned at roughly a 301 level: it assumes you have worked with the Claude API, built at least one agentic system, and have hands-on familiarity with MCP and Claude Code. Pure theory is not sufficient.

Specifically, the exam rewards:

1. Direct experience with the Messages API request-response cycle, including tool result appending.
2. Familiarity with Claude Code's configuration files and how they interact with CI environments.
3. Practical knowledge of MCP server setup, including environment variable expansion and scoping.
4. Experience debugging agentic loops, including premature termination and infinite loop anti-patterns.
5. Schema design for structured output, including how to prevent fabrication through field-level constraints.

Candidates who have only read documentation without building anything tend to struggle with the distractor options, which are designed to be plausible to someone who has not encountered the failure modes in practice.

How should you structure your preparation?

Given the domain weights, a rational study allocation looks like this:

We recommend starting with Domain 1 because its concepts underpin the other domains. An agent that mismanages its loop (D1) will also produce unreliable structured output (D4) and exhaust its context (D5). Understanding the architecture first makes the downstream domains easier to absorb.

Our concept library at /concepts covers 174 atomic concepts mapped to all five domains and 30 task statements. The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, so it will not advance you past a concept until you have demonstrated reliable recall. Practice exams are 60 questions, scored 100 to 1000 with 720 as the passing bar, matching the real exam format exactly.

AI Skill Certs is an independent prep platform and is not affiliated with, endorsed by, or approved by Anthropic.