6 Official CCAR-F Exam Scenarios
The CCAR-F exam does not test isolated syntax in a vacuum. Every question is anchored to one of six official production scenarios. Your exam will randomly draw four of these six scenarios.
How Anthropic Evaluates Scenario Architecture
In accordance with Anthropic Exam Guide v1.0, each candidate receives exactly 4 scenarios drawn at random from the canonical bank of 6. Each selected scenario anchors approximately 15 questions, distributing items across multiple blueprint domains.
Instead of asking "what is tool_choice?", items present real-world architectural dilemmas: "In Scenario 1, the customer support agent enters a continuous loop attempting to refund a closed account. Which configuration prevents this behavior while preserving idempotency?" Mastering all 6 scenario blueprints is the only reliable way to pass.
Architectural Teardown of All 6 Scenarios
Detailed breakdown of technical challenges, tested blueprint domains, key architectural patterns, and typical question traps for each scenario.
Customer Support Resolution Agent
Scenario ID: S1 • Tested Domains: Agentic Architecture & Orchestration, Tool Design & MCP Integration, Context Management & ReliabilityBuild an autonomous support resolution agent handling high-ambiguity customer inquiries (returns, billing disputes, account issues) using the Claude Agent SDK and custom MCP tools (get_customer, lookup_order, process_refund, escalate_to_human) targeting 80%+ first-contact resolution with deterministic escalation.
- Routing & classification with forced tool_choice
- MCP tool error isolation and customerSafeMessage separation
- Human-in-the-loop escalation calibration for refunds
Code Generation with Claude Code
Scenario ID: S2 • Tested Domains: Claude Code Configuration & Workflows, Context Management & ReliabilityAccelerate team software development workflows with custom slash commands (.claude/commands/), hierarchical CLAUDE.md configurations, path-scoped rules (.claude/rules/), and strategic plan mode vs direct execution decisions.
- CLAUDE.md memory hierarchy and project-level rule scoping
- Plan mode vs direct execution mode trade-offs
- Multi-file atomic refactoring and automated test execution
Multi-Agent Research System
Scenario ID: S3 • Tested Domains: Agentic Architecture & Orchestration, Tool Design & MCP Integration, Context Management & ReliabilityOrchestrate a hub-and-spoke multi-agent system where a coordinator agent delegates to specialized search, document analysis, synthesis, and reporting subagents with explicit context passing, provenance tracking, and error propagation.
- Orchestrator-subagent hierarchy with explicit task contracts
- Runaway loop circuit breakers and MAX_TURNS safety boundaries
- Shared memory synchronization and conflict resolution
Developer Productivity with Claude
Scenario ID: S4 • Tested Domains: Tool Design & MCP Integration, Claude Code Configuration & Workflows, Agentic Architecture & OrchestrationBuild developer productivity tools using the Claude Agent SDK to explore unfamiliar codebases, generate boilerplate, and automate tasks using built-in tools (Read, Write, Bash, Grep, Glob) and integrated MCP servers.
- JSON schema validation with Zod host-level retry loops
- Prompt caching optimization for repeated document templates
- Handling max_tokens truncation without schema invalidation
Claude Code for Continuous Integration
Scenario ID: S5 • Tested Domains: Claude Code Configuration & Workflows, Prompt Engineering & Structured OutputIntegrate non-interactive Claude Code into CI/CD pipelines to run automated code reviews, generate targeted test cases, enforce structured JSON output schemas, and minimize developer false-positive fatigue.
- Context window fact preservation and citation grounding
- Permission-boundary enforcement at the MCP tool layer
- Distinguishing genuine empty queries from access denial errors
Structured Data Extraction
Scenario ID: S6 • Tested Domains: Prompt Engineering & Structured Output, Context Management & ReliabilityExtract structured data from diverse, unstructured documents using tool_use with JSON schemas, schema validation-retry loops, nullable fields, and the Message Batches API for high-volume, non-blocking workloads.
- State machine transitions with deterministic condition guards
- Idempotent tool execution to survive network retry bursts
- Continuation turn prompting for incomplete model generations
How to Practice Scenario Decisions
1. Identify the Anchored Scenario
When reading an exam stem, immediately identify whether the system is a customer support agent, code generation pipeline, or multi-agent loop. The architectural constraints (latency, cost, security boundaries) depend on the scenario context.
2. Check Domain Boundaries
Most scenario questions require synthesizing at least two domains. For example, handling an unrecoverable MCP tool error (Domain 2) requires choosing the proper structured error schema (Domain 4) and deciding whether human escalation is required (Domain 5).
3. Reject Anti-Pattern Distractors
Distractors frequently suggest brute-force prompt additions ("tell Claude never to fail") or unconstrained loops. Correct architectural options enforce structural constraints via MCP schemas, hooks, and host verification loops.