Anthropic · CCAR-F · Foundations

Claude Certified Architect – Foundations (CCAR-F)

Also searched for as "CCA-F". The official exam code is CCAR-F.

Solution architects who design and build agent systems with Claude: agentic loops, multi-agent orchestration, tools and MCP, Claude Code configuration, structured output and reliability.

125 free practice questions · full timed mock exam · study plan · animated lessons

Exam facts

Checked 2026-11-15 against the official exam guide (v1.0, July 2026). Official exam guide · Official certification FAQ

Retakes, rescheduling and more

What the exam covers

D1 Agentic Architecture & Orchestration (27% of the exam)

D2 Tool Design & MCP Integration (18% of the exam)

D3 Claude Code Configuration & Workflows (20% of the exam)

D4 Prompt Engineering & Structured Output (20% of the exam)

D5 Context Management & Reliability (15% of the exam)

Scenarios (4 of these 6 appear on your exam)

Sample practice questions

D1 · scenario S1

A support agent built with the Agent SDK uses MCP tools get_customer, lookup_order, process_refund and escalate_to_human, targeting 80% first-contact resolution. The agent occasionally issues refunds before verifying identity, despite a clear rule in the prompt. What is the most reliable fix?

  • Restate the verification rule in capitals at the top and bottom of the system prompt
  • Move the agent to a larger model that follows multi-step instructions more reliably
  • Add ten few-shot conversations that always show verification happening before any refund
  • Have process_refund reject calls until verification has succeeded in this session

D2 · scenario S1

A support agent built with the Agent SDK uses MCP tools get_customer, lookup_order, process_refund and escalate_to_human, targeting 80% first-contact resolution. process_refund takes an order_id and an amount. Twice this month, a customer disputed one item and the agent refunded the whole order. What is the best change?

  • Lower the temperature so the agent picks refund amounts more consistently
  • Add a system prompt rule telling the agent to refund only the disputed items
  • Require the agent to call lookup_order right before every process_refund call
  • Make the tool take line-item IDs and compute the amount itself, rejecting others

D5 · scenario S1

A support agent built with the Agent SDK uses MCP tools get_customer, lookup_order, process_refund and escalate_to_human, targeting 80% first-contact resolution. A customer with a simple, solvable issue writes "I want to speak to a person." What should the agent do?

  • Escalate right away, passing a handoff summary that holds all the case facts
  • Ask the customer why they prefer a human so the agent can address that concern
  • Explain that the agent can handle this faster and carry on with the resolution
  • Resolve the simple issue first, then offer a human only if it stays unresolved

Answer these and more, with every option explained