InertiaRSS Track and read blogs, news, and tech you care about
Read Original Open in InertiaRSS

Recommended Feeds

让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
月光博客
月光博客
A
About on SuperTechFans
爱范儿
爱范儿
大猫的无限游戏
大猫的无限游戏
S
SegmentFault 最新的问题
G
Google Developers Blog
博客园_首页
云风的 BLOG
云风的 BLOG
Engineering at Meta
Engineering at Meta
博客园 - 叶小钗
The Cloudflare Blog
博客园 - 三生石上(FineUI控件)
GbyAI
GbyAI
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
T
The Blog of Author Tim Ferriss
雷峰网
雷峰网
WordPress大学
WordPress大学
量子位
酷 壳 – CoolShell
酷 壳 – CoolShell

博客园 - iTech

7万星的AI交易框架:让大模型模拟投行多空辩论,自动做交易决策 71000颗星的AI交易团队:让大模型模拟投行分工,自动做交易决策 13400颗星的开源项目:输入一句话,AI全自动帮你做短视频 102颗星的沙盒:当AI学会自己写代码、跑测试、做部署 AI 技术日报 - 2026-05-08 29k 星的 PageIndex:不用向量数据库,靠推理就能做 RAG 每天花两小时刷信息?这个开源项目帮你全自动搞定 读源码像读小说?试了 DeepWiki 和 Zread,我再也不想裸读 GitHub 了 Matt Pocock 开源的这套 .claude 技能,为什么让工程师集体上头? Cursor Team Kit:Cursor 官方团队在用的 17 个 AI 工作流 AI 技术日报 - 2026-05-07 AI 技术日报 - 2026-05-06 AI 技术日报 - 2026-05-05 Anthropic CEO 说 12 个月内程序员要失业,我扒完他的底牌,发现事情没那么简单 把工程师的肌肉记忆装进 Claude Code,这个 4300 Star 的项目我后悔没早用 AI 技术日报 - 2026-05-04 AI 技术日报 - 2026-05-03 AI 技术日报 - 2026-05-02 六大 Agent 框架横评:谁支持 Skills?谁能自动创建 Agent?MCP 呢? Wechatsync:一个 Chrome 插件,一键把文章同步到 31 个平台 LangChain 开源了 Open SWE:Stripe、Ramp、Coinbase 内部都在造的编程 Agent Cockpit:把 Claude Code 从终端里搬出来,装进浏览器 Cursor 把自家的 AI Agent 开放了:写几行 TypeScript 就能调 Cursor 干活 AI 技术日报 - 2026-05-01 AI 写代码每次结果都不一样?Archon 用 YAML 工作流把 AI 编程变成流水线 AI 写代码比你快了,但你还是得学编程——只不过学法得换 腾讯的龙虾特工队:4 个 AI Agent 同日更新,全家桶正式成型 Agno 不做更聪明的 Agent,它要把所有 Agent 框架包进同一个操作系统 Hermes Agent 终于有了像样的 Web 界面,而且还支持远程访问 Datawhale 出了一套 29 学科知识地图,把 AI 的底牌全掀了 Hermes Agent 在聊天框里就能用的 20 种高级功能 一份 AGENTS.md 能顶一次模型升级?Augment Code 用数据说了算 NVIDIA 开源了一个「AI 沙箱」,20K Star,让 Agent 跑代码不再裸奔 60ms 冷启动、5MB 内存:腾讯开源的这个沙箱让 Docker 安全隔离像笑话 AI 技术日报 - 2026-04-30 AI 技术日报 - 2026-04-29 AI 技术日报 - 2026-04-28 Goose:Linux 基金会亲儿子,能撼动 Claude Code 和 OpenCode 吗? AI 技术日报 - 2026-04-27 AI 技术日报 - 2026-04-26 Google 把价值20美元/月的东西免费了,102K人已经抢到了 OpenClaw 和 Claude Code 网络搜索配置指南 AI 技术日报 - 2026-04-25 Anthropic 为什么遥遥领先:从 Cat Wu 专访看AI霸主的底层逻辑 Mac 本地跑大模型完全指南:你的苹果电脑就是 AI 工作站 同样 70B 参数,为什么 MoE 只激活 13B 就能打平 Dense? DeepSeek-V4 技术报告里藏着一条线:华为昇腾 NPU 已完成推理验证 DeepSeek-V4 深夜炸场:1M 上下文、384K 输出、双模型,API 定价直接卷到底 MacBook Air 跑大模型实测:Ollama、llama.cpp、LM Studio 谁才是本地推理之王? AI 技术日报 - 2026-04-24
Anthropic launches the Claude Certified Architect exam: It's not about Prompt, but about whether you can build a production-level Agent
iTech · 2026-06-19 · via 博客园 - iTech

Finally, someone issued a "driver's license" to the AI engineer

In March 2026, Anthropic officially launched Claude Certified Architect – Foundations (CCA-F) Certification exam.

This is not an introductory course on "How to Use ChatGPT to Write Email". CCA-F is a serious architect level certification-60 situational multiple-choice questions, 120-minute time limit, invigilating environment, and a pass score of 720/1000. It tests not whether you can write a prompt, but whether you can use Claude to design, build and operate aproduction-levelAI system.

The exam is currently open to employees of Anthropic partner companies, and the first 5,000 partner employees take it for free.

What does the exam look like?

basic parameters

project details
Exam Name Claude Certified Architect – Foundations (CCA-F)
Question creator Anthropic official
question type 60 scene-based multiple-choice questions
duration 120 minutes
passing score 720 / 1000
examination form Online invigilation (Proctor)
scene Randomly select 4 scenarios from 6 scenarios
costs Free for the first 5,000 partner employees

You can't pass by reciting the question

Each question is embedded in a real production scene. You are not answering "What are the APIs of the Agent SDK?" You are answering:

"You are the architect of a customer support team. You need to design an Agent that can process refunds, check order status, and automatically upgrade to manual in complex situations. Which of the following architectures best fits Anthropic's best practices?"

You need to understand: when to use the hooks of the Agent SDK for compliance checks, when to use the MCP tool, when to let the Agent decide for itself, and when to require human approval.

5 major knowledge areas covering the entire ecosystem of Claude

The exam covers 5 domains, each with clear weights:

Domain 1:Agentic Architecture & Orchestration(27%)

This is the heaviest piece. The test is whether you can use Claude's Agent SDK to design and orchestrate a multi-agent system.

Core test points:
- Agentic Loops:循环的终止条件不是靠固定次数,而是靠 stop_reasontool_use vs end_turn
- Multi-Agent Orchestration:Hub-and-spoke 架构、coordinator-subagent 模式
- Lifecycle Hooks:用代码而不是 prompt 来强制执行业务规则
- Session Management:会话状态的恢复、分叉和隔离
- Task Decomposition:把一个大任务拆成多个子 Agent 并行执行

Domain 2:Tool Design & MCP Integration(18%)

考的是你怎么给 Agent 设计工具,以及怎么集成 MCP(Model Context Protocol)服务。

核心考点:
- 工具描述的最佳实践(description 写得不好 = 模型用不对)
- 结构化错误响应: Must containisError,errorCategory,isRetryableand context information_
_ - _Tool quantity control: Each Agent maintains 4-5 tools, and the model selection accuracy of more than 18 tools is significantly reduced
- Configuration and integration of MCP Server_
- Selection strategy for Claude's built-in tools (Read, Write, Bash, Grep, Glob)_

_Domain3: Claude Code Configuration_&_ Workflows (20%)

The test is how you configure Claude Code to support team development workflows.

Core test points:
- CLAUDE.md hier arch y: Global → Project → Directory → File Configuration Inheritance
- Custom slash commands and Skills
- Plan Mode vs Direct Execution: When to ask Claude to plan first and then implement it, and when to work directly
- Iterative optimization techniques
- CI/CD integration: using -p 标志进行非交互式运行、--output-format json 获取结构化输出

Domain 4:Prompt Engineering & Structured Output(20%)

不是「写个好 prompt」,而是怎么在生产系统中可靠地获得结构化输出。

核心考点:
- tool_use 模式:用 tool_use 而不是纯文本解析来获取结构化数据
- JSON Schema 设计
- Validation-Retry Loop:验证输出 → 不符合 schema → 自动重试
- Few-shot prompting 保持一致性
- Multi-instance & multi-pass review:用不同的 session 做 review,避免 reasoning context 偏见

Domain 5:Context Management & Reliability(15%)

The exam is the most easily overturned question in the production environment: context management.

Core test points:
- Risks of long context processing and progressive summary
- Context Transfer between Multiple Agents
- Error propagation: How do errors spread in multi-agent systems
- Escalation patterns: Upgrade based on task complexity and strategy gaps, not based on sentiment analysis
- Information traceability (Provenance) and multi-source synthesis
- Confidence calibration: Don't rely on the self-reported confidence of the model, use structured criteria to judge

6 exam scenarios (4 randomly selected)

The exam is not an abstract question and answer about knowledge points, but rather puts you in one of six real scenarios:

1. Customer Support Resolution Agent

Design a customer service Agent: handle consultations, solve problems, and upgrade complex situations manually. Test Agent SDK, MCP tools, and upgrade logic.

2. Code Generation with Claude Code

为开发团队配置 Claude Code 工作流。考 CLAUDE.md 配置、plan mode、slash commands、TDD 迭代。

3. Multi-Agent Research System

构建一个 coordinator-subagent 并行研究系统。考 multi-agent 编排、上下文隔离和传递、错误传播、结果综合。

4. Developer Productivity with Claude

用 Claude Agent SDK 构建开发者工具。考内置工具选择、MCP server 集成、代码库探索策略。

5. Claude Code for CI/CD

把 Claude Code 集成到 CI/CD 流水线。考 -p Non-interactive mode, structured output, Batch API, session isolation.

Extract structured data from unstructured documents. Test JSON schema design, tool_use, validation-retry loop, few-shot prompting.

The "Trap Answer" in the Exam

CCA-F's multiple-choice questions are very tricky-wrong options seem reasonable, but they are all common anti-patterns. Officials have summarized some high-frequency traps:

loop terminates

  • wrong: Use natural language parsing to decide whether to continue the cycle
  • on: Inspection stop_reasontool_usemeans that the tool still needs to be called,_end_turnmeans completed)_

  • Error: Set any iteration upper limit (For example,"Run up to 10 times") As the main stopping mechanism_

  • to: Let the agency loop pass throughstop_reasonNatural termination_

Business rule execution_

  • Error_: Use the prompt command to enforce critical business rules_
  • to: Use programmatic hooks (such asonToolCall) to make deterministic checks

Upgrade policy

  • wrong: Let the model report the confidence score itself, and then upgrade based on this
  • on: Use structured criteria and procedural checks to decide whether to upgrade

  • wrong: Upgrade based on sentiment analysis (user gets angry → switch to manual)

  • on: Upgrade based on task complexity and strategic gaps

error handling

  • wrong: General error message ("Operation failed")
  • on: Contains isErrorerrorCategoryisRetryable structural errors in context

  • wrong: Silent swallow error (empty result = success)

  • on: Distinguish between "access without permission" and "really no data"

tool management

  • wrong: Give an Agent 18+ tools
  • on: Each Agent maintains 4-5 tools, and Tool distribution will be done later

code review

  • wrong: Ask the Agent to generate code in the same session and then review it by itself_
  • Right to_: Use different sessions to make generators and reviewers, Avoid reasoning context bias_

Test preparation path_

Official free courses_

_Anthropic's certification platform (anthropiccertifications.com) provides a complete preparation course:

Claude Certified Architect -Foundations_(30 lessons)_
_ - Covering all 5 domains
_ - Includes animated illustrations, exam trap analysis, hands-on exercises, scenario quiz

Claude Code 101 (20 lessons)
- Claude Code from beginner to advanced
- Hooks, MCP Server, SDK Integration

Introduction to MCP (21 Lessons)
- Model Context Protocol End-to - End
- Architecture, primitives, build server and client

Introduction to Subagents(18 节课)
- 子 Agent 委托
- 上下文隔离、设计模式、多 Agent 编排

Introduction to Agent Skills(16 节课)
- 端到端构建 Agent Skills
- SKILL.md、progressive disclosure、多文件技能、分享

备考平台功能

  • Knowledge Graph:150+ 概念节点,可视化知识关联
  • Adaptive Practice: 500+ adaptive questions, strengthened for weaknesses
  • Mock Exams: 5 complete mock exams (60 questions/ 120 minutes)
  • FSRS Spaced Repetition: Scientific interval review schedule
  • AI Tutor: Claude-driven personalized tutoring

12-week study plan

The officially recommended 12-week structured study plan allocates the study progress of 5 domains on a weekly basis.

Test preparation fee

project costs
Certification platform basic edition Free (concept library, knowledge map, courses, interval review, adaptive exercises)
Certification platform Premium <$999 (one-time, approximately ¥85): AI Tutor, 5 sets of mock exams, performance dashboard
examination fee Free for the first 5,000 partner employees
CertSafari question bank 614 exercise questions (paid)

Who should take this test?

suitable person

  • Solution Architect: Using Claude API to design production-level AI applications
  • AI engineers: Responsible for the architecture and operation and maintenance of the Agent system
  • Platform Engineer: Promote Claude Code and MCP integration among teams
  • Technical Director: Need to assess the quality and safety of AI systems

wrong people

  • Beginners who just want to learn prompt engineering (take Anthropic's AI Fluency course)
  • Not in Anthropic Partner Company (temporarily unable to take the exam)
  • I just want to get a certificate to hang on LinkedIn (this exam really requires understanding of Agent architecture)

What does this certification say?

To be honest, certifications in the AI field currently do not have the same industry recognition as AWS Solutions Architect or Kubernetes CKA. The value of CCA-F depends on:

  1. Anthropic's market position: Claude's adoption rate in enterprise-level Agent development
  2. examination quality: From the perspective of scene design and anti-pattern analysis, this is not a certificate of "paying for it"
  3. actual needs: More and more companies are hiring engineers in the AI Agent direction, and an in-depth certification can be used as a screening criterion

But at least, the CCA-F exam content is designed seriously. It doesn't test the trivia of "How many parameters does Claude have?", but it tests "If your Agent goes wrong in the production environment, how can you make the system handle it gracefully instead of crashing"-this is what people who have actually done Agent development will ask.

how to register

  1. access Anthropic Skilljar registration page
  2. Submit an access request (currently limited to partner company employees only)
  3. After receiving exam access, take the exam online
  4. Test results will be scored immediately

in a word

CCA-F is not testing whether you know what Claude is, but whether you canUse it to build a system that actually works online--A system that can handle exceptions at 3 a.m., not crash when context explodes, and not lose data when multi-agent collaboration.

If this is what you do every day, this certification is worth a try. If you're preparing to do this, its test preparation material itself is an excellent guide to Agent architecture.


author: itech001
source: Public Account: AI Artificial Intelligence Era
website: _ _ JHSNS _ URL _ 0 _ _
Share the most cutting-edge AI news and technical research every day.

This article was first published in the era of AI artificial intelligence. Please indicate the source for reprinting.

参考来源:
- Claude Certified Architect — Foundations Course | Anthropic Certifications
- CCA-F Exam Guide | Claude Certifications
- CCA-F | Panaversity Certification
- Claude Certified Architect Foundations Training | Simpliaxis