## What is the CCA-F Certification? The Claude Certified Architect Foundations (CCA-F) is Anthropic's first official technical credential. Launched on March 12, 2026, the certification validates that you can design, build, and operate production-grade systems on top of the Claude API, Claude Code, the Claude Agent SDK, and the Model Context Protocol (MCP). Unlike vendor exams that test rote API recall, CCA-F is scenario-based. It puts you inside six realistic production scenarios -- customer support agents, multi-agent research pipelines, CI/CD automation, structured extraction systems, long-running coding agents, and tool-augmented workflows -- and asks you to make architecture decisions under real constraints. If you already build with Claude in production, the exam will feel like a thorough self-audit. If you are new to agentic systems, expect a steep but rewarding climb. ### Quick exam facts | Item | Detail | |------|--------| | Full name | Claude Certified Architect -- Foundations | | Code | CCA-F | | Issuer | Anthropic | | Format | Multiple choice, proctored | | Questions | 60 | | Duration | 120 minutes | | Passing score | 720 / 1000 (scaled) | | Fee | USD 99 (free for the first 5,000 Claude Partner Network employees) | | Delivery | Anthropic Academy on Skilljar | | Launch date | March 12, 2026 | ## Who Should Take CCA-F? CCA-F is positioned as a foundation-level credential, but "foundations" here means "foundations of building production agents," not "foundations of AI in general." The exam assumes you can read code, reason about system design, and understand HTTP, JSON schemas, and basic LLM concepts. The certification is a strong fit for: - **Software engineers** integrating Claude into existing products via the API - **AI/ML engineers** designing multi-agent or tool-using systems - **Solutions architects** evaluating Claude for enterprise workloads - **DevOps and platform engineers** wiring Claude Code into CI/CD - **Technical founders** building Claude-native applications - **Consultants and partners** delivering Claude implementations to clients It is not the right exam for absolute beginners with no programming background, or for prompt-engineering generalists who never touch the API. ## CCA-F Exam Domains and Weighting Anthropic publishes the official domain split. Memorise these percentages -- they tell you where to spend your study hours. | Domain | Weight | |--------|--------| | Agentic Architecture and Orchestration | 27% | | Claude Code Configuration and Workflows | 20% | | Prompt Engineering and Structured Output | 20% | | Tool Design and MCP Integration | 18% | | Context Management and Reliability | 15% | ### Domain 1: Agentic Architecture and Orchestration (27%) The largest domain, and the one that catches the most candidates out. You need to confidently reason about: - The **agentic loop** -- model call, tool execution, observation, repeat -- and where to break it - **Multi-agent patterns**: hub-and-spoke, orchestrator-worker, peer-to-peer handoff - **Lifecycle hooks**: `PreToolUse`, `PostToolUse`, `Stop`, `SubagentStop`, `UserPromptSubmit`, `SessionStart`, `SessionEnd`, `PreCompact`, `Notification` - **Subagents** in both Claude Code and the Agent SDK, including permission inheritance and per-subagent cost tracking - **Session management**, resumption, and compaction - **Task decomposition** strategies that survive context limits Anthropic's Agent SDK hooks reference is the authoritative source. See https://docs.claude.com/en/docs/agent-sdk/hooks for the full lifecycle map. ### Domain 2: Claude Code Configuration and Workflows (20%) This is the most configuration-heavy domain. Either you know where the files live or you do not. Topics include: - The `CLAUDE.md` file: project, user, and enterprise scope; import syntax (`@path/to/file.md`) - **Agent Skills** -- definition, discovery, parameters, and progressive disclosure - **Plan mode** and when to require explicit approval - **Slash commands** and custom command authoring - **Settings hierarchy**: `.claude/settings.json` vs `settings.local.json` vs user-level vs enterprise - **Hooks configuration** in settings (PreToolUse matchers, command vs prompt hooks) - **MCP server configuration** via `.mcp.json` and `${CLAUDE_PLUGIN_ROOT}` - CI/CD integration patterns and headless mode ### Domain 3: Prompt Engineering and Structured Output (20%) This is where wrong answers sound like good engineering. Be deliberate about: - **Context engineering**: what to load, when to load it, and what to omit - **Prompt caching**: cache breakpoints, TTL, the four-block cache hierarchy (tools, system, messages, conversation), `cache_control` placement - **Structured output** via JSON schemas and tool-based extraction - **Few-shot prompting** and example selection - **Vision** inputs and image-aware prompts - **Citations** for retrieval-grounded responses - **Extended thinking** vs standard responses and when each is worth the spend The canonical reference is https://docs.claude.com/en/docs/build-with-claude/prompt-caching. ### Domain 4: Tool Design and MCP Integration (18%) Underrated by candidates and over-tested by Anthropic. Focus on: - **Tool descriptions** as the primary lever for tool selection accuracy - **Structured error responses** that the model can recover from - **Tool boundaries** -- what belongs in one tool vs many - **MCP transports**: stdio, HTTP, SSE, WebSocket, and when each fits - **MCP servers**: resources, prompts, tools, sampling - **Tool search and `defer_loading`** to keep large tool catalogues out of the context window - **Built-in tools**: web search, code execution, computer use Authoritative docs: https://docs.claude.com/en/docs/agent-sdk/mcp. ### Domain 5: Context Management and Reliability (15%) Smallest weighting, but failures here cascade into every other domain. - **Long-context** strategies for 200K+ token windows - **Multi-agent handoffs** without losing state - **Error propagation** and structured failure modes - **Escalation paths** -- when to hand off to a human, a stronger model, or a different agent - **Token accounting** and cost ceilings - **Model selection**: Claude Opus 4.7 (`claude-opus-4-7`) for complex reasoning, Claude Sonnet 4.6 for balanced workloads, Claude Haiku 4.5 for high-throughput, latency-sensitive tasks ## Prerequisites Anthropic does not enforce hard prerequisites for CCA-F. There is no required prior certification and no mandatory training. That said, you will struggle without: - Comfort with at least one language that has an official Claude SDK (Python or TypeScript) - Familiarity with HTTP APIs, JSON, and environment variables - Hands-on time with Claude Code -- ideally 20+ hours of real project work - Basic understanding of agents, tools, and LLM cost structures If you have shipped one non-trivial Claude integration end-to-end, you have the right baseline. ## Exam Cost, Registration, and Retake Policy - **Cost**: USD 99 per attempt. Free for the first 5,000 Claude Partner Network employees through a dedicated voucher pool. - **Registration**: Anthropic Academy on Skilljar. You will need an Anthropic account and a government-issued ID for proctoring. - **Delivery**: Online proctored. - **Retake**: A 14-day waiting period applies between attempts. Each attempt requires a new fee unless your organisation has bulk vouchers. - **Validity**: Two years from the date you pass. Recertification will roll into future Foundations-level exam updates. ## How Scoring Works CCA-F uses a scaled score from 100 to 1000. You need 720 to pass. The raw-to-scaled conversion is not published, but community reports from candidates who passed suggest: - You can miss roughly 12 to 15 of the 60 questions and still clear 720 - Scenario sets often share a setup -- get the setup wrong and you can lose three or four marks in one go - There is no penalty for guessing, so never leave a question blank Results are delivered immediately at the end of the exam, with a domain-level breakdown on your score report. Your digital badge issues within 48 hours on Credly. ## Recommended Study Resources Start with primary sources, then layer in practice and community discussion. ### Official Anthropic resources - **Anthropic Academy preparation courses** -- 13 free, self-paced modules on Skilljar that map directly to the exam domains - **Claude API Docs** at https://docs.claude.com - **Claude Code Docs** at https://docs.claude.com/en/docs/claude-code - **Agent SDK reference** at https://docs.claude.com/en/docs/agent-sdk - **MCP specification** at https://modelcontextprotocol.io - **Anthropic Engineering blog** -- especially the posts on multi-agent research, code execution with MCP, and tool design ### Hands-on practice - Build at least one MCP server from scratch in either Python or TypeScript - Wire up a Claude Code project with a non-trivial `CLAUDE.md`, two or three custom skills, and a `PreToolUse` hook - Ship a small multi-agent system using the Agent SDK with explicit subagent boundaries - Instrument prompt caching on a real workload and measure the cache hit rate ### Practice exams Nothing replaces realistic, exam-format practice questions written against the published domains. PrepMyCert maintains a dedicated CCA-F practice exam course with full-length tests, detailed explanations, and domain analytics: https://www.prepmycert.com/course/ccaf-claude-certified-architect-foundations. ## A 6-Week Study Plan This plan assumes you have basic Claude API experience and can put in 8 to 10 hours per week. Adjust the slope to your starting point. ### Week 1 -- Foundations and Models - Read the Claude model overview. Memorise the three current models -- Opus 4.7, Sonnet 4.6, Haiku 4.5 -- and their positioning - Read the Messages API and tool-use reference end to end - Build a tiny script that calls the API with one tool and prints token usage - Read https://docs.claude.com/en/docs/build-with-claude/prompt-caching twice ### Week 2 -- Agent SDK and Hooks - Work through the Agent SDK quickstart in Python or TypeScript - Implement an agent with at least three custom tools - Add a `PreToolUse` hook that blocks one command and logs another - Read the hooks reference at https://docs.claude.com/en/docs/agent-sdk/hooks and note every event name ### Week 3 -- Claude Code Deep Dive - Configure a real project with a tiered `CLAUDE.md` (project + user + an imported file via `@path`) - Author two custom Agent Skills and one slash command - Set up a `.claude/settings.json` with a hook that runs lint on `PostToolUse` for Edit - Read the Claude Code plugin documentation and try installing a community plugin ### Week 4 -- MCP and Tool Design - Build an MCP server from scratch that exposes one resource, one prompt, and two tools - Connect it to Claude Code via `.mcp.json` and to an Agent SDK app - Experiment with `defer_loading` and tool search on a tool catalogue of 20+ tools - Practise writing tool descriptions that are unambiguous to the model ### Week 5 -- Multi-Agent Patterns and Context - Implement a hub-and-spoke pattern with two specialised subagents and one orchestrator - Force a context overflow on purpose and observe compaction behaviour - Add a `SubagentStop` hook that records per-subagent token cost - Write a structured-extraction agent using JSON schema and tool-based output ### Week 6 -- Practice, Review, and Sit - Take two full-length CCA-F practice exams under timed conditions - Review every wrong answer against the official docs -- no shortcuts - Rebuild any weak-domain example from scratch without looking at notes - Schedule the exam for the end of the week and rest the day before If you are short on time, the highest-leverage compression is to keep Weeks 2, 3, and 4 intact and trim Weeks 1 and 5. ## Common Mistakes That Sink Candidates From reviewing dozens of post-exam debriefs and our own question bank, these mistakes show up over and over: 1. **Confusing Claude Code hooks with Agent SDK hooks.** They share names but have different payloads and registration mechanisms. Know which one a question is asking about. 2. **Treating prompt caching as automatic.** Caching requires explicit `cache_control` breakpoints, and reordering or editing earlier content invalidates everything after the change. 3. **Picking Opus 4.7 for everything.** The exam rewards correct model selection. Use Haiku 4.5 for cheap, fast classification; Sonnet 4.6 for default agent work; Opus 4.7 only when reasoning depth justifies the cost. 4. **Ignoring tool descriptions.** Two tools with overlapping descriptions are the single most common cause of wrong tool calls. Tighten the language. 5. **Building one giant agent.** When a scenario describes mixed responsibilities, the right answer is usually subagents with narrow scope, not a single super-prompt. 6. **Forgetting that subagents do not inherit permissions.** Each spawned subagent re-requests tool permissions. Plan for this in your settings. 7. **Reading scenarios too fast.** Each scenario sets constraints (latency, cost, compliance) that change the right answer. Read twice. 8. **Skipping MCP transports.** Know stdio vs HTTP vs SSE vs WebSocket and which fits a given deployment. ## Frequently Asked Questions ### How hard is the CCA-F exam? Candidates with one or more shipped Claude integrations typically pass on the first attempt with three to four weeks of focused prep. Candidates without hands-on experience report needing six to eight weeks and at least one practice-exam cycle. ### How much does the CCA-F exam cost? USD 99 per attempt. The first 5,000 Claude Partner Network employees receive a free voucher. Some employers reimburse the fee under their professional development budget. ### Is CCA-F worth it in 2026? If you build with Claude professionally, yes. It is the only first-party credential, it is recognised by the Claude Partner Network, and it materially sharpens your production architecture skills as a side effect of preparing. ### How long is the CCA-F certification valid? Two years from the date you pass. Anthropic has signalled that recertification will likely be a shorter delta exam covering changes since your last pass. ### Can I take CCA-F without partner access? Yes. Anyone with an Anthropic account can register and pay the USD 99 fee. The Anthropic Academy preparation courses are free for everyone, partner or not. ### What programming language should I prepare in? Python and TypeScript are the two officially supported Agent SDK languages. Either is fine. The exam shows pseudo-code rather than language-specific syntax, but you will study faster in whichever language you use day to day. ### Are there official practice questions from Anthropic? Anthropic Academy includes a small set of sample questions in some modules, but there is no full-length official practice exam. PrepMyCert's CCA-F practice exam course fills that gap with realistic, scenario-format questions and detailed explanations. ### What is the difference between CCA-F and a future CCA-Pro? CCA-F is the Foundations tier. Anthropic has hinted at a Professional tier (CCA-P) covering deeper performance tuning, security, and large-scale multi-agent orchestration, but no firm date has been announced. Pass Foundations first. ### How many hours of hands-on practice do I actually need? To clear the bar comfortably, plan on around 40 to 60 hours of real keyboard time across the SDK, Claude Code, and MCP. Passive reading without building leaves blind spots that scenario questions expose ruthlessly. If you have less than 20 hands-on hours, prioritise building over re-reading docs. ### Is there a discount or voucher for CCA-F? The USD 99 fee is the standard rate. The first 5,000 Claude Partner Network employees receive a free voucher through their partner admin. Some employers reimburse the fee under professional development, and Anthropic has run occasional community vouchers for hackathon winners and Anthropic Academy completers. Watch your Anthropic Academy inbox. ## Exam-Day Tips That Actually Move the Needle Small, boring habits add up to a real score difference on exam day. From candidates who passed on the first attempt: - **Read each scenario twice before looking at the answer options.** Scenarios bury constraints in the third or fourth sentence. Latency budgets, compliance requirements, and cost ceilings all change which answer is right. - **Eliminate two answers before considering the remaining two.** Most CCA-F questions have one obvious distractor and one near-miss. Knock the obvious one out first. - **Watch for absolute language.** Words like *always*, *never*, *only*, and *every* in an answer choice are usually wrong. Production engineering rarely has absolutes. - **Flag and move.** If a question is taking more than three minutes, flag it and come back. You cannot afford to lose two easy questions later because you stalled on one hard one. - **Trust the published weighting.** When you are torn between two answers, pick the one that aligns with the largest-weighted domain principles -- usually orchestration and Claude Code configuration. - **Pre-exam logistics matter.** Online proctoring will fail your check-in for a cluttered desk, a second monitor that is not unplugged, or an ID with poor lighting. Test your setup the day before. ## Next Step The single highest-leverage thing you can do after reading this guide is to sit a realistic, timed practice exam and see where the gaps actually are. Reading docs creates the illusion of readiness; questions reveal the truth. When you are ready, take the CCA-F practice exam at https://www.prepmycert.com/course/ccaf-claude-certified-architect-foundations. Full-length, domain-weighted, and updated as Anthropic's official blueprint evolves. Good luck on exam day. Read every scenario twice.