Claude Certified Architect, Foundations practice questions
Solution architects designing production Claude systems
What the CCAR-F exam covers
July 2026 guideSample CCAR-F questions
Five of the 60 practice questions, each with its answer and explanation.
Agentic architecture and orchestration
In an agentic loop built on the Messages API, which signal should decide whether the loop continues?
- Whether the reply text contains the word 'done'
- Whether stop_reason is 'tool_use' (continue) or 'end_turn' (finish)
- A fixed count of exactly three iterations
- Whether the reply is longer than 500 tokens
Show the answer
B. Whether stop_reason is 'tool_use' (continue) or 'end_turn' (finish)
The stop reason is the structured signal. Parsing the text or using an arbitrary iteration count as the main control is unreliable.
Agentic architecture and orchestration
Policy says refunds above $500 must never be issued without a manager's approval. Where should this rule be enforced?
- In the system prompt only
- In a programmatic check or hook that blocks the refund tool call until approval is recorded
- In a few-shot example
- In the tool's name
Show the answer
B. In a programmatic check or hook that blocks the refund tool call until approval is recorded
Prompt instructions are followed most of the time, not every time. A rule that must always hold needs deterministic enforcement in code.
Agentic architecture and orchestration
A coordinator agent delegates research to a subagent, and the result ignores constraints discussed earlier in the conversation. What is the most likely cause?
- The subagent's model is too small
- The subagent does not inherit the coordinator's conversation, so the constraints were never passed to it
- Subagents cannot use tools
- The coordinator's temperature is too high
Show the answer
B. The subagent does not inherit the coordinator's conversation, so the constraints were never passed to it
Subagents start with their own context. Everything they need must be included explicitly in the task they are given.
Tool design and MCP integration
An MCP tool fails because the order ID does not exist. Which response helps the agent recover best?
- An empty success result
- A structured error flagged as an error, saying what went wrong and whether a retry could help
- The text 'Error' with no detail
- A crash with no response
Show the answer
B. A structured error flagged as an error, saying what went wrong and whether a retry could help
A structured error lets the agent tell a bad input from a temporary fault and choose between correcting, retrying or escalating.
Tool design and MCP integration
Your team wants everyone working in a repository to get the same MCP servers automatically. Where should the configuration live?
- In each developer's personal user-level settings
- In a project-level .mcp.json committed to the repository, with secrets supplied through environment variables
- In a chat message at the start of every session
- In the README only
Show the answer
B. In a project-level .mcp.json committed to the repository, with secrets supplied through environment variables
Project-scoped configuration travels with the repository. Personal or experimental servers belong in user-level settings, and secrets stay out of version control.