The complete open-source roadmap for learning AI Agents — from LLM basics to production-ready Agent systems.
Agent Learning (agent_learning) is a systematic, practice-oriented AI Agent learning roadmap and hands-on tutorial covering LLM fundamentals, RAG, memory, tool use, function calling, agentic workflows, LangChain, LangGraph, MCP, multi-agent systems, evaluation, deployment, and agentic RL.
If you want to learn how to build AI Agents — not just use ChatGPT, but understand how agents retrieve knowledge, remember context, call tools, plan actions, collaborate, and run safely in production — this project is for you.
Daily auto-tracking of arXiv frontier papers — content stays cutting-edge, always.
-4CAF50?style=for-the-badge){kind=icon}
-2196F3?style=for-the-badge){kind=icon}
Daily updated Agentic-RL frontier research chapter |
Step-by-step GRPO / GSPO learning content |
🤖 This repository automatically searches arXiv for the latest AI Agent-related papers every day and updates the content accordingly — ensuring you always stay at the cutting edge of research!
- 📡 Daily Automated Search: A scheduled pipeline scans arXiv daily for new papers on Agent architectures, tool use, memory systems, multi-agent collaboration, reinforcement learning for agents, and more.
- 📝 Auto-Updated Content: Relevant findings are automatically integrated into the corresponding chapters, keeping the book's frontier sections fresh and up-to-date.
- 🔔 Never Miss a Breakthrough: No need to manually track dozens of research feeds — this repo does it for you, so you can focus on learning and building.
💡 This means the content you read here is not static — it evolves continuously with the latest advances in the AI Agent field.
- Developers who want to build real AI Agent applications instead of only prompting chatbots
- Students and beginners who need a structured path from LLM basics to Agent systems
- LLM application engineers working with RAG, tool calling, memory, LangGraph, MCP, and evaluation
- Researchers and builders who want to connect frontier Agent papers with engineering practice
- Product and startup teams exploring production-ready Agent workflows
| Path | Start Here | Goal |
|---|---|---|
| Beginner Path | LLM basics → Prompt Engineering → Function Calling → RAG → Memory → ReAct | Understand how an Agent works end to end |
| Engineering Path | Tool Layer → LangGraph → Evaluation → Security → Deployment → Observability | Build production-ready Agent systems |
| Research Path | ReAct → Reflexion → MemGPT → PPO / DPO / GRPO → Agentic RL | Follow and understand frontier Agent research |
| Project Path | Hello Agent → RAG QA Agent → Memory Agent → Data Analysis Agent → Coding Agent | Learn by building complete applications |
- 🎯 Step by Step: From LLM fundamentals to multi-Agent systems, each chapter has a clear knowledge progression
- 💻 Code First: Every core concept comes with runnable Python code examples
- 🎨 Rich Illustrations: 120+ hand-drawn SVG architecture diagrams / flowcharts / sequence diagrams for intuitive understanding
- 🎬 Interactive Animations: 5 built-in interactive HTML animations (Perceive-Think-Act cycle, ReAct reasoning, Function Calling, RAG flow, GRPO sampling)
- 🔬 Paper Reviews: Key chapters include frontier paper deep-dives (ReAct, Reflexion, MemGPT, GRPO, etc.)
- 🏗️ Complete Projects: 3 comprehensive hands-on projects (AI Coding Assistant, Intelligent Data Analysis Agent, Multimodal Agent)
- 🛡️ Production Ready: Covers security, evaluation, deployment, and other production essentials
- 🧪 Cutting Edge: Covers Context Engineering, Agentic-RL (GRPO/DPO/PPO), MCP/A2A/ANP, and other 2025–2026 latest advances
- 📐 Formula Support: KaTeX-rendered math formulas for clear reading of policy gradient, KL divergence derivations in RL chapters
- 🔄 Continuously Updated: Tracking the latest changes in LangChain, LangGraph, MCP, and other frameworks
Below are selected showcases from the book's 120+ hand-drawn SVG illustrations, all original to this book.
|
Perceive-Think-Act Loop (Chapter 1) 
Agent's core mechanism: Perceive environment → LLM reasoning → Execute action → Loop until goal achieved |
ReAct Reasoning Framework (Chapter 5) 
Thought → Action → Observation alternating loop, enabling Agents to think while acting |
|
Function Calling Complete Flow (Chapter 3) 
6-step complete flow from user input to tool invocation to final response, with message structure illustration |
RAG Retrieval-Augmented Generation (Chapter 6) 
Offline indexing + Online retrieval dual-phase architecture, making LLM answers evidence-based |
|
Three-Layer Memory Architecture (Chapter 4) 
Working memory → Short-term memory → Long-term memory, with important info sinking down and semantic retrieval pulling up |
Prompt Engineering vs Context Engineering (Chapter 7) 
From "how to say it" to "what the LLM sees" — the paradigm shift of the Agent era |
|
Three Multi-Agent Communication Patterns (Chapter 15) 
Message Queue (async decoupling) / Shared Blackboard (data sharing) / Direct Call (real-time collaboration) |
MCP / A2A / ANP Protocol Comparison (Chapter 16) 
Three-layer protocol stack: ANP for discovery → A2A for task collaboration → MCP for tool invocation |
|
GRPO Training Architecture (Chapter 10) 
No Critic model needed, computes advantage via intra-group normalization, only 1.5× model size in VRAM |
LangGraph Three Core Concepts (Chapter 12) 
State (shared state) · Node (processing unit) · Edge (execution flow control) |
📖 The above is just a selected preview — For the full 120+ architecture diagrams + 5 interactive animations, please read online
This book includes 5 interactive HTML animations to help you intuitively understand the dynamic processes of core concepts:
| Animation | Chapter | Description |
|---|---|---|
| 🔄 Perceive-Think-Act Cycle | Chapter 1 | Dynamic demonstration of Agent's core loop |
| 💡 ReAct Reasoning Process | Chapter 5 | Shows the alternating Thought → Action → Observation process |
| 🔧 Function Calling | Chapter 3 | Complete tool invocation flow animation |
| 📚 RAG Retrieval Flow | Chapter 6 | From document chunking to vector retrieval to answer generation |
| 🎯 GRPO Sampling Process | Chapter 10 | Visualization of intra-group multi-output sampling and reward normalization |
💡 Interactive animations are only available in the online e-book. Local builds can also preview them.
|
🧠 Agent Core Architecture
🛠️ Tools & Skills
🧪 Reinforcement Learning Training
|
💾 Memory, Knowledge & Context
🤝 Multi-Agent Collaboration & Communication
🛡️ Production Full Pipeline
|
# Install mdBook (choose one)
cargo install mdbook
# Or macOS: brew install mdbook
# Install mdbook-katex plugin (for math formula rendering)
cargo install mdbook-katex
# Clone the repository
git clone https://github.com/Haozhe-Xing/agent_learning.git
cd agent_learning
# Build both Chinese and English versions and start unified server (default port 3000)
./serve.shAfter starting, visit:
- 🌐 Language Selection Home:
http://localhost:3000 - 🇨🇳 Chinese Version:
http://localhost:3000/zh/ - 🇺🇸 English Version:
http://localhost:3000/en/
# Python 3.11+
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# Install core dependencies
pip install langchain langchain-openai langgraph openai anthropic
# Configure API Key
export OPENAI_API_KEY="your-key-here"
-191919?style=flat){kind=icon}
All forms of contribution are welcome!
- 🐛 Found a bug: Submit an Issue
- 💡 Content suggestions: Start a Discussion
- 📝 Improve content: Fork → Edit → Submit PR
- ⭐ Support the project: Give this repo a Star!
# Fork and clone
git clone https://github.com/YOUR_USERNAME/agent_learning.git
# Create a feature branch
git checkout -b feature/improve-chapter-3
# Local preview
./serve.sh
# Commit and push
git commit -m "feat: improve Chapter 3 tool calling code examples"
git push origin feature/improve-chapter-3- Each chapter is placed in a separate directory
src/zh/chapter_xxx/(Chinese) orsrc/en/chapter_xxx/(English) - Chapter overview goes in
README.md, sections are numbered as01_xxx.md,02_xxx.md - Chinese SVG illustrations go in
src/zh/svg/, English versions insrc/en/svg/, naming format:chapter_xxx_description.svg - Chinese interactive animations go in
src/zh/animations/, English versions insrc/en/animations/
All paper reading and frontier research sections should follow a consistent structure so that readers can quickly understand why a paper matters, what it contributed at the time, and how it connects to real Agent engineering.
Use the following template for each representative paper:
### Paper Title: one-sentence explanation of the problem it solves
- **Paper link**:
- **Code / project link**:
- **Year / organization**:
- **Problem addressed at the time**:
- **Core contribution**:
- **Method breakdown**:
- **Engineering insight for Agent systems**:
- **Limitations**:Quality requirements:
- Link to the original source: include the arXiv, conference, official blog, GitHub, or project page whenever available.
- Explain historical contribution: describe what problem the work solved when it appeared, not only what it does.
- Connect to engineering practice: explain how the idea affects Agent memory, tools, planning, evaluation, safety, training, or deployment.
- State limitations: clarify what the paper does not solve, where assumptions are strong, or whether the result is mainly benchmark-driven.
- Avoid paper lists without synthesis: after several papers, add a short comparison table or narrative summary explaining how the works relate to each other.
This project is open-sourced under the MIT License.
- Chinese / English online book powered by mdBook
- 120+ original SVG architecture diagrams and flowcharts
- Interactive animations for core Agent concepts
- Paper reading sections for key Agent research
- Agentic RL chapters covering PPO / DPO / GRPO
- Runnable Agent example projects and templates
- Agent glossary and keyword cheat sheet
- Agent architecture diagram gallery
- Interview questions and self-check exercises
- Production-ready Agent template with evaluation and observability
If this project helps you, please give it a Star ⭐ — it's the greatest encouragement for the author!
Built with ❤️, so that every developer can master AI Agent development