Pass the Claude Certified Architect Foundations Exam (CCAR-F)

1. A support engineering team is building an autonomous ticket-triage agent on the Claude Agent SDK. Their logs show the agent calling Read, Grep, and a custom CRM lookup repeatedly, but the developers cannot articulate what actually ends a run. They want to instrument the end of each run so they can record cost and outcome.In the Agent SDK's agent loop, which condition causes the loop to terminate and a ResultMessage to be emitted under normal completion?

   • A. Claude produces a response that contains no tool call requests
   • B. The configured max_turns value is reached
   • C. Every tool listed in allowed_tools has been invoked at least once
   • D. The response carries stop_reason set to tool_use

   Answer: A. The loop is: receive prompt, evaluate and respond, execute tools, repeat. It ends when Claude returns output containing no tool calls, at which point the SDK yields the final AssistantMessage followed by a ResultMessage. B describes a limit that terminates the run abnormally with the error_max_turns subtype, not normal completion. C is invented; the agent calls only the tools it needs. D is backwards: stop_reason tool_use means Claude wants a tool result and the loop must continue.

   Exam domain: Agentic Architecture & Orchestration

2. During an architecture review, a product manager asks the architect to explain what the team's dashboards mean when they report that a run took 'nine turns'. The dashboard is fed from the SDK's message stream.Which description of a turn in the agent loop is accurate?

   • A. One user prompt and the single assistant reply it produces, regardless of tool activity
   • B. One round trip: Claude emits tool calls, the SDK runs them, results feed back
   • C. One HTTP request to the Messages API, including any streaming events it emits
   • D. One complete session from the init SystemMessage through to the ResultMessage

   Answer: B. A turn is one round trip through the loop: Claude emits tool calls, the SDK runs them, and the results return to Claude automatically without yielding control to your code. A describes a single conversational exchange, which can span many turns. C conflates a turn with an API call; a turn can involve several tool executions and the loop continues without new user input. D describes a whole session, which normally contains many turns.

   Exam domain: Agentic Architecture & Orchestration

3. A small startup wants an agent that can execute code and browse files, but has no platform team, no container orchestration, and no appetite for operating sandboxes. They want to call an HTTP endpoint and get agent behavior back.Which option matches this constraint?

   • A. Managed Agents, where Anthropic hosts both the agent and the sandbox
   • B. The Claude Agent SDK deployed on their own container platform
   • C. The Claude Code CLI run headlessly on a developer laptop as a scheduled job
   • D. The client SDK with a hand-written loop calling a self-hosted execution environment

   Answer: A. Managed Agents is the hosted REST option where Anthropic operates the agent loop and the sandbox, which is precisely the requirement. B still requires the team to run and scale the process. C puts production workload on an unmanaged laptop with no availability story. D is the most operational work of all, since they must build both the loop and the execution environment.

   Exam domain: Agentic Architecture & Orchestration

4. Developer Productivity with ClaudeA developer wants Claude Code to follow their personal commit-message style in every repository they work in on their own laptop, without adding anything to any project's version control. Where should that instruction live?

   • A. In `~/.claude/CLAUDE.md`
   • B. In a root `CLAUDE.md` committed to each repository they work in
   • C. In a `memory` key inside `.claude/settings.local.json`
   • D. In a managed policy `CLAUDE.md` installed by the organization's device management

   Answer: A. A is correct: the user-level memory file at `~/.claude/CLAUDE.md` loads in every session on that machine and lives outside any repository, which is exactly the stated requirement. B fails because a root `CLAUDE.md` is checked into version control and imposes one person's preference on the whole team, and it would have to be repeated per repository. C fails because `settings.json` files carry configuration keys such as permissions and hooks; memory instructions are not a settings key. D fails on scope and ownership: managed policy memory is administrator-controlled, applies to everyone on the fleet, and is not something an individual developer installs for a personal preference.

   Exam domain: Claude Code Configuration & Workflows

5. A developer needs to record their personal sandbox URL and a few local-only paths for one project, without those details reaching teammates through version control. Which file is intended for this?

   • A. `.claude/settings.json` at the project root
   • B. `~/.claude/CLAUDE.md` in the developer's home directory
   • C. A rule file at `.claude/rules/local.md`
   • D. `CLAUDE.local.md` at the project root, gitignored

   Answer: D. D is correct: `CLAUDE.local.md` exists for private, per-project preferences, loads alongside the project memory file, and is meant to be gitignored. A fails because the shared project settings file is checked in and is configuration, not memory. B fails on scope: a user-level memory file applies to every project on the machine, so project-specific sandbox details would leak into unrelated work. C fails because `.claude/rules/` is normally committed and shared with the team, which is the opposite of the requirement.

   Exam domain: Claude Code Configuration & Workflows

Want more? Finish the quiz above to unlock 15 more free questions, or open the full CCAR-F practice app.