首頁 / Home
Trilingual — the English edition is fully maintained, not a thin machine translation (only ~0.4% of English lines carry any CJK, almost all intentional bilingual term-mapping). zh-TW is the curation source of truth (new content lands there first); the English and 简中 editions track the same structure, with CI checking localization correctness and anchor integrity across all three.
Learning roadmap + 145+ curated resources + simple illustrative cases — three pillars helping you go from "I don't know where to start" to "I can design multi-agent systems". Structured 8-stage path from LLM fundamentals to multi-agent orchestration, Computer Use / Browser Use / Code Sandbox.
🎯 Why this exists¶
What this repo is: a learning roadmap + 145+ curated resources + simple illustrative cases — three pillars helping AI / AI-agent learners go from "I don't know where to start" to "I can design multi-agent systems."
Concretely:
| Pillar | What it does | Scale |
|---|---|---|
| Learning roadmap | Organizes scattered high-quality projects, tutorials, and required reading into 8 stages (including Stage 5 + Stage 8 as two shared hubs) + 2 tracks + 5 specialized branches, from zero to advanced | 8 stages, 2 tracks |
| Resource curation | Each stage curates 145+ projects (star rating, audience, what they teach, how to run) plus an MCP/Skill catalog covering the Chinese AI ecosystem (DeepSeek, Zhipu, Kimi, …) | 145+ projects, 62 MCP/Skill |
| Simple illustrative cases | Each stage ships 1-5 foundational exercises (70-150 line starter + dual-path Ollama/Anthropic SDK comparison + mock-based tests) | 23 exercise folders |
After the main path, you go from "LLM user" to "agent system builder" — capable of designing multi-agent collaboration, writing your own MCP server, and shipping real agent systems.
📋 Table of Contents¶
- 🎯 Why this exists
- 📚 Quick Start
- 🗺️ Learning Map (Two Tracks)
- 💡 How to Learn
- 📚 Related Resources
- 🤝 Contributing
- 🙏 Acknowledgments
- 🎓 Citation
- License
📚 Quick Start¶
🚀 First time with AI agents / never written code before?¶
Start here: resources/setup-guide.en.md — 30-45 minutes from zero, walks you through getting an API key, installing Python, and running your first LLM hello-world.
Read online¶
- Learning Map (Two Tracks) — read this section to decide Track A or Track B
- Stage 0 Foundations — already know Python / git / API? Skip straight to Stage 1
Local clone¶
git clone https://github.com/WenyuChiou/awesome-agentic-ai-zh.git
cd awesome-agentic-ai-zh
# Start with stages/00-foundations.en.md
✨ What you get¶
- 📖 Fully free — MIT-licensed, all content open
- 🗺️ Two learning tracks — Track A (CLI Power User) for "use existing CLIs"; Track B (Agent Builder) for "build your own". Shared Stages 0-2 foundation.
- 🛠️ Foundational hands-on exercises — 1-5 illustrative exercises per stage (specs + dual-path SDK comparison + success criteria). Positioned as foundational + roadmap verification — for chapter-length depth exercises see the hello-agents / Anthropic Cookbook callout in each stage
- 🎯 145+ curated projects — each with star rating, audience, what it teaches, how to run (incl. local LLM runners: Ollama, llama.cpp, LocalAI, MLX)
- 🌏 Trilingual, fully maintained — zh-TW (canonical) / 简中 / English; the English edition is complete, not a thin mirror
- 🎓 Beyond frameworks: Claude Code ecosystem — MCP / Skills / Plugins / SDK full stack
- 🔬 5 specialized branches — researcher / developer / teacher / knowledge worker / everyday user
- ⏱️ Time commitment, stated upfront — Track A 8-10 weeks / Track B 16-22 weeks minimum, 5-7 months realistic (5-8 hr/week part-time)
🗺️ Learning Map (Two Tracks)¶
After Stages 0-2 (shared foundations), pick a track based on your goal:
- Track A — CLI Power User: you want to USE existing CLI agents (Claude Code, Codex, OpenCode, Gemini CLI, etc.) to get work done — not build agents from scratch. 3 sub-stages (A1-A3).
- Track B — Agent Builder: you want to BUILD your own agents — learn frameworks, write ReAct, design multi-agent systems. Stages 3-7 main path.
The two tracks are not mutually exclusive — most people start with A to get hands-on, then come back to B for internals (or vice versa). Stage 5 (Claude Code Ecosystem) is used by both tracks.
Shared Foundations (Stages 0-2)¶
| Stage | Topic | Key Content | Time |
|---|---|---|---|
| 0 | Foundations | Python · CLI · git · API · JSON | 1-2 wks |
| 1 | LLM Fundamentals | tokens · API · model comparison · local LLM | 1 wk |
| 2 | Prompt Engineering | system prompts · few-shot · CoT | 1-2 wks |
Track A — CLI Power User (use CLIs to get work done)¶
| Stage | Topic | Key Content | Time |
|---|---|---|---|
| A1 | CLI Agent Intro & Selection | 7-CLI comparison · install · first run | 1 wk |
| A2 | CLI Workflow Patterns | CLAUDE.md · slash commands · multi-step decomposition | 1-2 wks |
| A3 | Integration & Production | MCP-into-CLI · CI automation · cost / observability | 1-2 wks |
| +5 | Stage 5 — Claude Code Ecosystem (Shared Hub) | MCP · Skills · Plugins · Subagents; Track A reads 5.1-5.4 (5.5-5.6 optional) | 1-2 wks (Track A view) |
| +8 | Stage 8 — Agent Interfaces (Shared Hub) | Computer Use · Browser Use · Code Sandbox; Track A reads Track A usage | 1-2 wks (Track A view) |
Track A total time: includes Stages 0-2 (shared foundations) + A1-A3 + Stage 5 + Stage 8 (two shared hubs) ≈ 8-10 weeks. Core reference:
resources/cli-agents-guide.en.md.
Track B — Agent Builder (build agents from scratch)¶
| Stage | Topic | Key Content | Time |
|---|---|---|---|
| 3 ⭐ | Tool Use & Hello Agent | function calling · ReAct · 5 hands-on exercises | 2-3 wks |
| 4 | Agent Frameworks | LangGraph · AutoGen · CrewAI · Smolagents | 2-3 wks |
| 5 ⭐⭐ | Claude Code Ecosystem (Shared Hub, Track A also studies) | MCP · Skills · Plugins · Subagents | 3-4 wks (Track B view) |
| 6 | Context Engineering: RAG and Memory | vector DB · long-term memory · contextual retrieval | 2 wks |
| 7 | Multi-Agent · Productionization | multi-agent orchestration · eval · observability · advanced SDK | 2-4 wks |
| 7.5 | Advanced Agentic Workflow Concepts (reading map) | work boundary · PAR loop · agent-as-judge · 12 advanced concepts + reading list | 1 wk (no code) |
| 8 ⭐⭐ | Agent Interfaces (Shared Hub, Track A also studies) | Computer Use · Browser Use · Code Sandbox; 2024-2026 frontier | 2-3 wks (Track B view) |
Track B total time: minimum 16-22 weeks, realistic 5-7 months (5-8 hr/week part-time)
Two shared hubs (used by both Track A + Track B): - Stage 5 = Claude Code Ecosystem (MCP / Skills / Plugins / Subagents) — Track A learns MCP-into-CLI, Track B learns agent runtime structure - Stage 8 = Agent Interfaces (Computer Use / Browser / Sandbox, 2024-2026 frontier) — Track A learns "how to use" for task delegation, Track B learns "how to build" with embedded interfaces
💡 Want a concrete cross-stage example? Build Your First AI Agent in 7 Steps — same Paper Summary Bot traced from Stage 1 through Stage 7, ~350 lines of executable code (Track B)
After the main path, pick one of 5 specialized branches. Not sure which?
💡 The Everyday User branch can be read directly without walking the main path — it's for people who want to use AI without writing code.
| Branch | Best for | Topics |
|---|---|---|
| 🔬 Researcher | Grad students, postdocs, PIs | Lit triage · paper writing · multi-agent review |
| 💻 Developer | Software engineers | Cursor · Aider · CLI delegation · code review |
| 🎓 Teacher | Teachers, instructors | Lesson planning · slides · student feedback · privacy / ethics · prompt templates |
| 📊 Knowledge Worker | Consultants, PMs, analysts | Email · meeting notes · report automation |
| 👥 Everyday User | ChatGPT / Claude.ai users | Daily writing · learning · privacy · CLI agent intro |
💡 How to Learn¶
Welcome — future agent system builder. Some guidance before you start.
This roadmap balances concepts with hands-on work, helping you transform from an LLM user into an agent system builder. It assumes basic Python. Before starting:
- Basic Python — written functions, used APIs, can read JSON
- Basic git — clone, commit, push
- Motivation to learn — agents are the fastest-changing area in AI 2025+, and require sustained effort
If anything's missing, do Stage 0; if not, start at Stage 1.
The main path has 5 parts:
- Part 1 (Stages 0-2): Foundations & LLM Basics — Python / git / API, what's an LLM, prompt design
- Part 2 (Stages 3-4): Build Your Agent — from tool use to agents, learn the major frameworks
- Part 3 (Stage 5) Shared Hub — Claude Code Ecosystem (MCP / Skills / Plugins / Subagents; used by both Track A + B)
- Part 4 (Stages 6-7): Advanced Integration — memory / RAG / multi-agent collaboration / harness engineering
- Part 5 (Stage 8) Shared Hub — Agent Interfaces (Computer Use / Browser Use / Code Sandbox, 2024-2026 frontier; used by both tracks)
🔭 Three layers of concept evolution: prompt engineering (Stage 2 — how to write a single prompt) → context engineering (Stage 3 onward — how to dynamically assemble system prompt + memory + retrieved chunks + tool schema) → harness engineering (Stage 7 — agent loop / eval / observability / deploy as a complete production system). Three terms, three phases; you don't need to look elsewhere. See
stages/02-prompt-engineering.en.md"Beyond prompts: context engineering" andstages/07-multi-agent-production.en.mdRequired Reading 5+6.
After the main path (16-22 weeks for Track B, 8-10 weeks for Track A), pick a branch.
The most important advice: don't skip the hands-on exercises. Each stage's exercises are "you can't learn this without doing it" — skim past them and you'll get stuck later.
🎓 How to actually use the exercises: the
starter.pyin each exercise folder is a complete solution, not a TODO skeleton. If you clone,cat starter.py, and runpython test.pyto all-green, you'll think "I learned it" — but you haven't written a single line. Correct learning loop:mv starter.py starter_reference.py, look at signatures (not bodies), write your own, peek at the reference only after 20 min stuck. Full method + per-stage time budgets + escalation order indocs/HOW_TO_USE.md.
Ready? Start at Stage 0.
📚 Related Resources¶
The full related-resources block (term definitions + daily-tool MCP/Skill highlights + awesome lists + Chinese-community resources) lives in RESOURCES.en.md so this README stays focused.
Common quick links, grouped by scenario:
🚀 Onboarding / Environment¶
| Your situation | Where | What's there |
|---|---|---|
| Never written code, first time with AI agents | resources/setup-guide.en.md |
30-45 min from zero (API key, Python, first hello-world) |
| Not sure which LLM provider to pick | resources/setup-guide.en.md A |
Anthropic / OpenAI / DeepSeek / Kimi / NVIDIA NIM comparison |
| Topic-based awesome lists / Chinese community | RESOURCES.en.md topic-based |
5-10 min skim |
📖 Concepts / Terminology¶
| Your situation | Where | What's there |
|---|---|---|
| Don't know a term (LLM / agent / RAG / token / MCP / Skill / vector DB…) | resources/glossary.en.md |
30+ terms, 30-80 words each + which stage covers it |
| Why some agents live in terminal vs Telegram vs Jetson | resources/agent-paradigms.en.md |
5 paradigms mental model + Hermes Agent / OpenClaw examples |
| MCP / Skills / Plugins glossary mapping | RESOURCES.en.md three core terms |
1-page lookup |
🛠 Hands-on¶
| Your situation | Where | What's there |
|---|---|---|
| Want to build Skill / MCP server / Word / Zotero / local LLM integration | resources/cookbook.en.md |
6 step-by-step recipes, 30-50 min each |
| Want to use subagents but do not know who to dispatch, how to dispatch, or what work to dispatch | resources/subagent-cookbook.en.md |
15 copy-paste dispatch recipes |
| Stuck on tool calling (LLM won't call / schema broken / ReAct won't stop) | examples/stage-5/tool-calling-tutor/ |
Claude Code installable skill, 4-symptom diagnostic |
| How to use the hands-on exercises correctly (active vs passive mode) | docs/HOW_TO_USE.md |
5-10 min read, applies to every stage |
🔌 Daily tool integrations / Finding MCP servers¶
| Your situation | Where | Scope |
|---|---|---|
| Connect to Notion / Obsidian / Excel / GitHub / etc. | RESOURCES.en.md daily-tool integrations |
7-8 highlights |
| Full MCP server / Skill catalog (stars, categories) | resources/mcp-skills-catalog.en.md |
62 entries, 6 categories |
🔬 Research / Production¶
| Your situation | Where | What's there |
|---|---|---|
| Research workflow + multi-LLM delegation skill pair | RESOURCES.en.md research workflow |
Maintainer's own Claude Code research skill set |
| CLI agent 7-way comparison + production combos | resources/cli-agents-guide.en.md |
Track A's core reference, ~148 lines |
| Schema design rules (must-read for tool calling) | resources/schema-design-cheatsheet.en.md |
5 golden rules + 5 anti-patterns |
🤝 Contributing¶
This repo is an AI learning document — if you've also curated great resources, contributions are very welcome:
- 🐛 Bug reports — wrong content, broken links, stale info → open Issue
- 💡 Suggestions — missing stage / new project to add → open Issue to discuss
- 📝 Improvements — refine existing stage content, fix typos → direct PR
- ✍️ Add a project — 1-3 new projects per stage with "why this teaches that stage" rationale
- 🌏 Translations — improve the English edition or translate to other languages
- 🌱 Become a Stage / Branch maintainer — long-term review of a specific area, see CONTRIBUTORS.md
PR process and style rules: CONTRIBUTING.md + resources/style-guide.en.md.
📅 Want to see what shipped recently? →
CHANGELOG.md(last 14 days). Internal phase rollout progress and launch checklist:.github/launch-checklist.md(maintainer-facing internal doc).
🙏 Acknowledgments¶
Inspiration¶
- Datawhale Hello-Agents — the most thorough chapter-length agent tutorial in the Chinese-language ecosystem; inspired our chapter + progress structure. Every stage / exercise folder has a 📚 callout pointing to the relevant depth chapter. Special thanks.
- Datawhale community — landmark Chinese ML learning community; multiple anchor projects come from them
- liyupi/ai-guide — largest Chinese-language "AI mega-guide" + Vibe Coding tutorial (covers Agent Skills / RAG / MCP / A2A / Harness Engineering). This repo is a "structured roadmap"; ai-guide is a "breadth resource hub" — complementary
Related projects¶
Other lists in the same space — useful to browse alongside this repo when hunting for specific tools:
wong2/awesome-mcp-servers— categorized MCP server catalogpunkpeye/awesome-mcp-servers— another MCP server cataloghesreallyhim/awesome-claude-code— Claude Code tools & plugins list
These are pure catalogs (browse and pick). This repo is different in that it has a learning order from Stage 0 all the way to production.
Contributors¶
New contributors appear above automatically. Full list → GitHub Contributors.
Personal¶
- @WenyuChiou — Maintainer
🎓 Citation¶
If this learning roadmap helps your study or work, please cite:
@misc{awesome_agentic_ai_zh_2026,
title = {awesome-agentic-ai-zh: A Structured Learning Roadmap for Agentic AI},
author = {Chiou, Wenyu},
year = {2026},
url = {https://github.com/WenyuChiou/awesome-agentic-ai-zh},
note = {8-stage learning path from prerequisites to Agent Interfaces (Computer Use / Browser Use / Code Sandbox), with curated projects + hello-X demos. Bilingual (zh-TW / English).}
}
📈 Star History¶
License¶
MIT. Maintained by @WenyuChiou.
⭐ If this repo helps you, please give it a Star — it matters for ongoing iteration