Claude Certified Architect - Foundations (CCA-F): what it is and what it tests
Anthropic's foundational certification for building production AI agents with Claude. The five domains, the exam format, the kind of questions it asks, and how to prepare.
Most AI certifications test whether you can describe machine learning. The Claude Certified Architect - Foundations (CCA-F) tests something more practical and more current: whether you can build a real agentic system with Claude - one that uses tools, manages context, and behaves reliably in production. It's the first certification on CertLabPro from an AI-model provider rather than a cloud provider, and it reflects where a lot of 2026 engineering work has actually moved: designing agents, not just calling a model.
Here's what the cert is, exactly what it covers, the kind of questions it asks, and how to prepare.
What CCA-F is
CCA-F is a foundational-level certification focused on the architecture and engineering of Claude-based applications and agents. "Foundations" here doesn't mean trivia - it means the core competencies every Claude builder needs before specializing: how to structure an agent loop, how to design tools, how to write prompts that produce reliable structured output, how to use Claude Code, and how to keep long-running agents stable.
If you've been building with Claude - wiring up tool use, the Model Context Protocol (MCP), multi-agent setups, or Claude Code workflows - this cert validates that knowledge. If you're newer to it, the blueprint below is a good map of what "knowing Claude" actually means in practice.
Exam format
| Code | CCA-F |
| Provider | Anthropic |
| Category | AI |
| Level | Foundational |
| Questions (exam) | 60 |
| Time | 120 minutes |
| Passing score | 720 / 1000 (≈72%) |
| Practice questions on CertLabPro | 255 |
The questions are scenario-based, not definitional. You're rarely asked "what is X" - you're asked "given this situation, which design is correct," which is why the practice bank (255 questions) matters more than memorizing terms.
The five domains
The exam is weighted across five domains. This is the part to study against:
| # | Domain | Weight |
|---|---|---|
| 1 | Agentic Architecture & Orchestration | 27% |
| 2 | Claude Code Configuration & Workflows | 20% |
| 3 | Prompt Engineering & Structured Output | 20% |
| 4 | Tool Design & MCP Integration | 18% |
| 5 | Context Management & Reliability | 15% |
1. Agentic Architecture & Orchestration (27%)
The biggest domain. How to structure agents: the reason-act-observe (ReAct) loop, when to use a single agent vs. multiple, how an orchestrator should delegate to sub-agents, and where the safety boundaries go. Expect questions about choosing the right pattern for a task and about the risks of giving an orchestrator unrestricted tool access.
2. Claude Code Configuration & Workflows (20%)
Working with Claude Code as an agentic coding environment - how it's configured and driven, and the workflow patterns that make it productive (skills, hooks, sub-agents, and structured task execution).
3. Prompt Engineering & Structured Output (20%)
Getting Claude to produce reliable, machine-usable output: system prompts, when and how to constrain output to a schema, and prompting techniques that improve consistency for downstream automation.
4. Tool Design & MCP Integration (18%)
Designing tools an agent can actually use well - clear names and descriptions, the right granularity, and connecting capabilities through the Model Context Protocol (MCP). Good tool design is one of the highest-leverage skills in agent building, and the exam treats it that way.
5. Context Management & Reliability (15%)
Keeping long-running and multi-turn agents stable: managing conversation history and context, persistent memory, and reliability guardrails like maximum-iteration limits and graceful fallbacks when an agent gets stuck.
What the questions are actually like
The CCA-F question bank is built around real engineering decisions. A few representative examples (paraphrased):
- An agent must research a topic, write code, run tests, and iterate until they pass - which agentic pattern fits? (Answer: a ReAct loop with tool use and observation-based iteration.)
- A customer-support agent needs to look up orders, process refunds, escalate to a human, and hold multi-turn context - what architecture is right? (Answer: an agentic loop with a tool per capability, conversation-history management, and a system prompt defining role and escalation rules - not a stateless function, not fine-tuning, not RAG alone.)
- An orchestrator delegates to specialized sub-agents - what's the risk of giving it unrestricted tool access? (Answer: it may bypass the sub-agents and act directly, breaking the separation of concerns and safety boundaries.)
- An agentic coding assistant sometimes loops forever when tests keep failing - best mitigation? (Answer: a maximum-iteration count plus a fallback that asks the user for guidance.)
Notice the pattern: every question is a design trade-off. Knowing the concepts (ReAct, MCP, structured output, context windows) isn't enough - you need to know which one to reach for in a given scenario, and why the alternatives are wrong.
Who should take it
- Developers building with Claude - agents, tool use, MCP servers, or Claude Code - who want a credential that maps to what they already do.
- Engineers moving into agentic AI from general software or cloud roles, who want a structured way to learn the patterns rather than picking them up ad hoc.
- Teams standardizing on Claude that want a shared baseline for "what good looks like" in agent design.
It pairs naturally with cloud AI foundations like AWS AI Practitioner (AIF-C01) or Azure AI Fundamentals (AI-900): those cover the broad AI/ML landscape, while CCA-F goes deep on the building-with-an-agentic-model skills those don't.
How to prepare on CertLabPro
CertLabPro covers CCA-F with 255 practice questions mirroring the five domains and the scenario-based style of the exam, plus the same study modes as every other cert - Practice, Exam simulation, Flashcards, Weakest Link, SRS, and more - so you can drill the domains where you're weakest and rehearse under timed, 60-question conditions before the real thing. There's also a hands-on lab for the Anthropic track if you want to learn by doing rather than only by answering questions.
The bottom line
CCA-F is a focused, practical credential for the agentic era: it certifies that you can architect and operate Claude-based agents, not just talk about AI. The five domains - agentic architecture, Claude Code, prompting and structured output, tool/MCP design, and context/reliability - are the actual skill set behind shipping agents that work. If that's the direction your work is going, it's a credential worth having, and the fastest way to it is drilling the scenarios until the right design is obvious.