Claude Certified Architect – Foundations (CCAR-F)
Study site for the Claude Certified Architect – Foundations certification, built against Exam Guide v1.0 (effective July 2026).
The credential validates that a practitioner can reason about tradeoffs when building production systems with Claude across four surfaces: Claude Code, the Claude Agent SDK, the Claude API, and the Model Context Protocol (MCP in Claude Code).
Items are scenario-based and drawn from real customer use cases: agentic customer support, multi-agent research pipelines, Claude Code in CI/CD, developer productivity tooling, and structured extraction from unstructured documents. Conceptual recall alone is not sufficient — items test architectural judgment.
Exam details at a glance
| Attribute | Value |
|---|---|
| Credential | Claude Certified Architect – Foundations |
| Exam code | CCAR-F |
| Number of items | 60 |
| Item format | Multiple-choice and multiple-response; each item states how many responses to select |
| Exam structure | 4 scenarios drawn from a bank of 6 |
| Time limit | 120 minutes (~2 min/item) |
| Delivery | Proctored — online proctored and/or Pearson VUE test center |
| Passing score | Scaled score of 720 on a 100–1,000 scale |
| Exam fee | $125 USD |
| Validity | 12 months from the award date |
| Result reporting | Pass/fail with scaled score, plus percent-correct by domain |
Blueprint and domain weights
| # | Content domain | Weight | Approx. items | Task statements |
|---|---|---|---|---|
| 1 | Agentic Architecture & Orchestration | 27% | ~16 | 7 |
| 2 | Tool Design & MCP Integration | 18% | ~11 | 5 |
| 3 | Claude Code Configuration & Workflows | 20% | ~12 | 6 |
| 4 | Prompt Engineering & Structured Output | 20% | ~12 | 6 |
| 5 | Context Management & Reliability | 15% | ~9 | 6 |
| Total | 100% | 60 | 30 |
Domain 1 is the largest single block; Domains 3 and 4 tie for second. Domains 1, 2 and 5 dominate the agentic scenarios, while 3 and 4 dominate the Claude Code and extraction scenarios.
Intended audience
The target candidate is a solution architect with 6+ months of hands-on production experience with Claude. Assumed experience:
- Building agentic applications with the Claude Agent SDK — multi-agent orchestration, subagent delegation, tool integration, lifecycle hooks
- Configuring Claude Code for team workflows —
CLAUDE.mdfiles, Agent Skills, MCP server integrations, plan mode - Designing MCP tool and resource interfaces for backend system integration
- Engineering prompts for reliable structured output — JSON schemas, few-shot examples, extraction patterns
- Managing context windows across long documents, multi-turn conversations, and multi-agent handoffs
- Integrating Claude into CI/CD for automated code review, test generation, and PR feedback
- Making escalation and reliability decisions — error handling, human-in-the-loop, self-evaluation
The five decision rules
Nearly every item is a "which change is most effective" question. Five rules resolve the majority of them.
| Situation | Answer |
|---|---|
| A rule must hold every time — identity verification before a refund, a refund cap, a required tool ordering | Programmatic gate — a hook or prerequisite check. Never a stronger system prompt: prompt compliance is probabilistic. |
| The model picks the wrong tool among similar tools | Expand the tool descriptions — purpose, inputs, boundaries, worked examples. Not few-shot, not a routing classifier. |
| The judgment boundary is fuzzy — escalate vs resolve, flag vs skip | Explicit criteria plus 2–4 few-shot examples. Not a classifier, not self-reported confidence, not sentiment. |
| Output is inconsistent or contradictory across many files | Decompose into focused passes. A larger context window does not fix attention dilution or position effects. |
| Something crosses an agent or process boundary — errors, handoffs, subagent output, extraction results | Structured metadata — category, retryability, source, date, confidence, partial results — so the consumer can decide rather than guess. |
Underneath all five: fix the root cause at the smallest effective layer. Bad tool selection caused by thin descriptions is fixed by better descriptions, not by a routing classifier. Prefer the low-effort, high-leverage change before adding infrastructure.
Distractor smells: train a classifier · self-reported confidence as a gate · sentiment thresholds · a bigger model or a bigger context window · blaming a downstream component that is working correctly · consensus voting across full-scope runs · flags that do not exist.
What to build before the exam
Items test architectural judgment, so the item writers assume you have built these systems. Reading is not sufficient.
| Focus | What to actually do | Domains |
|---|---|---|
| Build an agent with the Claude Agent SDK | Implement a complete agentic loop with tool calling, error handling, and session management. Spawn subagents and pass context between them. | 1, 2, 5 |
| Configure Claude Code end to end | CLAUDE.md at project and directory level, path-specific rules in .claude/rules/, custom skills with context: fork and allowed-tools, and at least one MCP server. | 3, 2 |
| Design and test MCP tools | Write descriptions that clearly differentiate similar tools — see Writing effective tools for agents. Implement structured error responses with categories and retryable flags. Test selection reliability with deliberately ambiguous requests. | 2 |
| Build a structured extraction pipeline | tool_use with JSON schemas, validation-retry loops, optional/nullable fields, and the Message Batches API. | 4 |
| Practise prompt engineering | Few-shot examples for ambiguous scenarios. Explicit review criteria to cut false positives. Multi-pass review architectures for large reviews. | 4 |
| Study context management | Extract structured facts from verbose tool outputs, implement scratchpad files for long sessions, design subagent delegation against context limits. See Effective context engineering for AI agents. | 5 |
| Review escalation and HITL | When to escalate (policy gaps, customer requests, inability to progress) vs resolve autonomously. Human review workflows with confidence-based routing. | 5 |
Four preparation exercises turn that list into buildable projects. Each one sits on the domain page it reinforces:
| Exercise | Build | Page |
|---|---|---|
| 1 | Multi-tool agent with escalation logic | 1.1 Agentic loops |
| 2 | Claude Code for a team development workflow | 3.1 CLAUDE.md hierarchy |
| 3 | Structured data extraction pipeline | 4.3 Structured output |
| 4 | Multi-agent research pipeline | 1.2 Coordinator–subagent orchestration |