Claude Certified Architect – Foundations (CCAR-F)

CCAR-F exam facts

Every row comes from Anthropic's exam guide and Partner Academy FAQ. Scaled means the score is equated across exam forms, so it doesn't convert to a percentage.

Source: Claude Certified Architect – Foundations Exam Guide, version 1.0, effective July 2026 and the Partner Academy certification FAQ, accessed September 25, 2026.

What's on the CCAR-F exam?

The exam covers 5 domains, weighted by Anthropic's job task analysis. Agentic Architecture & Orchestration is the biggest block at 27% of scored questions. Exam items are written against these 30 task statements:

The scenarios behind the questions

Each sitting shows 4 of these 6 scenarios at random, and every question hangs off one of them. Prepare for all of them:

1. Customer Support Resolution Agent
2. Code Generation with Claude Code
3. Multi-Agent Research System
4. Developer Productivity with Claude
5. Claude Code for Continuous Integration
6. Structured Data Extraction

What's not on the exam

The guide rules these out, so skip them when you're short on time:

• Fine-tuning or training custom models
• API authentication, billing and account management
• Deploying or hosting MCP servers
• Computer use and vision
• Streaming and server-sent events
• Rate limits, quotas and API pricing
• Cloud-provider-specific configuration (AWS, GCP, Azure)
• Prompt caching internals and tokenization

A 4-week CCAR-F study plan

This plan assumes you already build with Claude at work. It follows the domain weights, heaviest first, and turns the guide's own preparation list into weekly builds:

1. Week 1: agentic architecture (27%)

   Build one complete agent with the Claude Agent SDK. Drive the loop off stop_reason, append tool results to the history, spawn a subagent with explicit context, and block a risky tool call with a hook. Then resume and fork a session.

2. Week 2: tools, MCP and Claude Code (18% + 20%)

   Write tool descriptions that keep two similar tools apart and return structured MCP errors with a retryable flag. Add a project-scoped MCP server. In Claude Code, set up a CLAUDE.md hierarchy, path-scoped rules and a skill, and run one non-interactive review with JSON output.

3. Week 3: prompts, structured output and context (20% + 15%)

   Extract data with tool_use and a JSON schema, then add a validation-retry loop. Use few-shot examples to cut false positives and try the Message Batches API. Rehearse the escalation rules and trim long tool output before it floods the context.

4. Week 4: timed practice

   Sit full practice exams at about 2 minutes a question. Sort your misses by domain, reread those task statements below, and drill the weakest domain until it stops being the weakest.

Sample CCAR-F practice questions

Anthropic's exam content is confidential under the candidate NDA, so none of these are exam questions. They come from SecuSpark's own architect practice bank, which is SecuSpark-generated with AI assistance, one question per domain.

1. Q1SecuSpark practice question (unofficial) · Agentic Architecture & Orchestration

   An insurance carrier is designing a multi-agent assistant to investigate disputed claim denials. A coordinator can delegate to a policy-language analyst, a claim-timeline analyst, and a regulatory analyst. The architecture team wants the system to be observable and to handle partial subagent failures consistently. Which communication pattern best fits this requirement?

   1. A.Allow the policy-language analyst to directly call the claim-timeline analyst whenever it needs chronology details, then send a final summary to the coordinator.
   2. B.Use a hub-and-spoke pattern where the coordinator invokes subagents, routes all inter-subagent information, handles errors, and aggregates results.
   3. C.Run all three subagents independently and let the regulatory analyst decide which outputs are relevant for the final answer.
   4. D.Give every subagent the complete conversation transcript and shared memory so they can negotiate responsibilities among themselves.

   Show the answer

   Answer: B. In a coordinator-subagent architecture, the coordinator should manage communication, error handling, routing, and result aggregation rather than letting subagents coordinate independently.

2. Q2SecuSpark practice question (unofficial) · Tool Design & MCP Integration

   A data governance group has an MCP server that can query warehouse tables for a Claude Code workflow. Analysts complain that Claude repeatedly calls tools to list tables and inspect schemas before answering basic questions about available datasets. The group wants Claude to see what data exists without making exploratory tool calls each time. What should the architects add?

   1. A.Move the MCP server from project-scoped .mcp.json to user-scoped ~/.claude.json so each analyst can configure only the tables they care about.
   2. B.Add a run_sql tool description that says Claude should call it whenever it needs to know which tables exist.
   3. C.Expose the warehouse schema catalog as MCP resources, such as database schemas and documented table hierarchies.
   4. D.Rely on MCP connection-time discovery because all configured server tools become available simultaneously to the agent.

   Show the answer

   Answer: C. MCP resources can expose content catalogs, giving agents visibility into available data and reducing exploratory tool calls.

3. Q3SecuSpark practice question (unofficial) · Claude Code Configuration & Workflows

   A developer productivity team wants Claude Code to review pull requests (PRs) in a continuous integration/continuous delivery (CI/CD) workflow and post findings as inline comments. Their comment-posting script expects machine-parseable JavaScript Object Notation (JSON) matching fields such as file, line, severity, and message. Which approach best supports this requirement?

   1. A.Ask Claude Code to write findings in Markdown tables because they are easier for humans to inspect in CI logs.
   2. B.Save the review output to a plain text artifact and have a separate model invocation infer the file and line numbers later.
   3. C.Run Claude Code interactively during PR validation and require reviewers to copy the findings into the PR UI.
   4. D.Invoke Claude Code with -p, --output-format json, and --json-schema that defines the expected finding fields.

   Show the answer

   Answer: D. CI review output should be non-interactive and structured using JSON output plus a schema for machine parsing.

4. Q4SecuSpark practice question (unofficial) · Prompt Engineering & Structured Output

   A software-as-a-service company wants Claude to perform license-risk checks on pull requests before developers can merge. The policy requires the check result to be available in the continuous integration (CI) gate within 10 minutes. Leadership asks whether the team should switch to the Message Batches API to reduce cost. What should the architect recommend?

   1. A.Keep this workflow on the synchronous Claude API because the check blocks merging and has a strict latency requirement.
   2. B.Switch to the Message Batches API because its 50% cost savings make it the preferred option for all repeated CI checks.
   3. C.Use the Message Batches API but submit one batch per pull request so each request remains isolated and easier to monitor.
   4. D.Run the pre-merge checks through batch overnight and allow developers to merge immediately, then revert any risky changes detected later.

   Show the answer

   Answer: A. Blocking pre-merge checks require predictable responsiveness, while batch processing has up to a 24-hour window and no guaranteed latency service-level agreement.

5. Q5SecuSpark practice question (unofficial) · Context Management & Reliability

   A property-management company uses a Claude-powered assistant to handle tenant maintenance requests. A tenant writes, "I do not want to troubleshoot this with a bot. Transfer me to a human agent now." The issue appears straightforward: the tenant only needs to reschedule an existing repair appointment. What should the assistant do?

   1. A.Acknowledge the request, then try to complete the rescheduling because the task is within the assistant’s capability.
   2. B.Run the appointment lookup tool first, and escalate only if the tool returns an error or conflicting appointment records.
   3. C.Immediately escalate to a human agent without first attempting to resolve the appointment change.
   4. D.Use sentiment analysis to decide whether the tenant is upset enough to warrant escalation.

   Show the answer

   Answer: C. An explicit customer request for a human is an escalation trigger that should be honored immediately, even if the issue seems simple.