# agent-study

**Repository Path**: m626/agent-study

## Basic Information

- **Project Name**: agent-study
- **Description**: AI Agent全栈课程：从ReAct循环到Claude Code逆向、MCP/A2A协议、RAG、DSPy、生产可观测性——全部为可运行Python文件，面试导向。
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-07
- **Last Updated**: 2026-06-07

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

<p align="center">
  <h1 align="center">🤖 AI Agent Full-Stack Learning Course</h1>
  <p align="center">
    Master AI Agent core theory and engineering from zero to one<br>
    36 Chapters · 22,000+ Lines of Code · 60+ Runnable Examples · Interview-Ready
  </p>
</p>

<p align="center">
  <img src="https://img.shields.io/badge/Python-3.10+-blue?logo=python&logoColor=white" alt="Python">
  <img src="https://img.shields.io/badge/Chapters-36-green" alt="Chapters">
  <img src="https://img.shields.io/badge/License-MIT-yellow" alt="License">
  <img src="https://img.shields.io/badge/Updated-2026.05-brightgreen" alt="Update">

  <br>

  <a href="README.md">🇨🇳 中文 README</a>
</p>

---

## 🔗 Quick Links

<p align="center">
  <a href="https://agent-study-ruddy.vercel.app/"><img src="https://img.shields.io/badge/🌐-Live_Site-Vercel?style=for-the-badge&logo=vercel&logoColor=white&labelColor=000000&color=6366F1" alt="Live Site"></a>
  <a href="https://github.com/Callous-0923/agent-study"><img src="https://img.shields.io/badge/⭐-GitHub_Repo-181717?style=for-the-badge&logo=github&logoColor=white" alt="GitHub"></a>
  <a href="https://agent-study-ruddy.vercel.app/chapter_09_rag_deepdive/09_rag_deepdive.html"><img src="https://img.shields.io/badge/🔍-RAG_Deep_Dive-FF6B6B?style=for-the-badge" alt="RAG Deep Dive"></a>
  <a href="https://agent-study-ruddy.vercel.app/chapter_10_mcp/10_mcp_deepdive.html"><img src="https://img.shields.io/badge/🔌-MCP_Protocol-10B981?style=for-the-badge" alt="MCP Protocol"></a>
</p>

> **🌐 No installation — read in your browser**: [agent-study-ruddy.vercel.app](https://agent-study-ruddy.vercel.app/) · All 36 chapters · Dark blueprint theme · Mobile-friendly

---

## 📖 Overview

A **career-oriented** AI Agent full-stack learning course. From Agent fundamentals to Claude Code reverse engineering, from RAG to MCP/A2A protocols, from DSPy to production observability — covering **36 topics across 7 tiers**. Each chapter is a **standalone runnable `.py` file** — a complete lecture and executable code in one.

> **Target audience**: New graduates, career switchers, and any developer looking to systematically master AI Agents.

---

## 🗺️ Course Roadmap (36 Chapters · 7 Tiers)

```
Tier 1: Foundations ── Tier 2: Engineering ── Tier 3: Deep Tech ── Tier 4: Production
   Ch0-3                   Ch4-7                  Ch8-12               Ch13-18

Tier 5: Advanced Arch ── Tier 6: Reinforcement ── Tier 7: Expert
  Ch19-24                    Ch25-28                   Ch29-36
```

### Tier 1: Agent Foundations

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch0](https://callous-0923.github.io/agent-study/chapter_00_overview/00_course_overview.html)** | Course Overview & Setup | Learning roadmap, dependency installation, API key config |
| **[Ch1](https://callous-0923.github.io/agent-study/chapter_01_fundamentals/01_hello_agent.html)** | Your First Agent | Handwritten ReAct loop, Function Calling principles |
| **[Ch2](https://callous-0923.github.io/agent-study/chapter_02_components/02_agent_components.html)** | Agent Core Components | Planner, memory systems (short/long/working term), tool design golden rules |
| **[Ch3](https://callous-0923.github.io/agent-study/chapter_03_types/03_agent_types.html)** | Agent Architecture Types | ReAct / Plan-Execute / Reflexion comparison |

### Tier 2: Engineering Practice & Frameworks

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch4](https://callous-0923.github.io/agent-study/chapter_04_frameworks/04_frameworks.html)** | Mainstream Frameworks | LangChain Agent + LangGraph state machine |
| **[Ch5](https://callous-0923.github.io/agent-study/chapter_05_multi_agent/05_multi_agent.html)** | Multi-Agent Systems | Writer+Reviewer collaboration, crewAI style |
| **[Ch6](https://callous-0923.github.io/agent-study/chapter_06_evaluation/06_evaluation.html)** | Evaluation & Testing | Eval frameworks, LLM-as-Judge, production checklist |
| **[Ch7](https://callous-0923.github.io/agent-study/chapter_07_interview/07_interview_prep.html)** | Interview Preparation | 20 high-frequency questions + project guide + interview flow |

### Tier 3: Deep Technical Analysis

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch8](https://callous-0923.github.io/agent-study/chapter_08_claude_code/08_claude_code_architecture.html)** | Claude Code Architecture | nO main loop, h2A Steering, context compaction, SubAgent |
| **[Ch9](https://callous-0923.github.io/agent-study/chapter_09_rag_deepdive/09_rag_deepdive.html)** | RAG Deep Dive | Chunking strategies, Embedding selection, RRF, Cross-Encoder, production modes |
| **[Ch10](https://callous-0923.github.io/agent-study/chapter_10_mcp/10_mcp_deepdive.html)** | MCP Protocol Explained | JSON-RPC, primitives (Tools/Resources/Prompts), capability negotiation |
| **[Ch11](https://callous-0923.github.io/agent-study/chapter_11_tool_calling/11_tool_calling_deepdive.html)** | Tool Calling Internals | OpenAI vs Anthropic, streaming assembly, strict mode |
| **[Ch12](https://callous-0923.github.io/agent-study/chapter_12_infrastructure/12_infrastructure.html)** | Agent Production Infrastructure | OpenClaw architecture, Harness, production checklist |

### Tier 4: Production Engineering

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch13](https://callous-0923.github.io/agent-study/chapter_13_fastapi/13_fastapi_agent_service.html)** | FastAPI Agent Service | REST API, SSE streaming, WebSocket, production deployment |
| **[Ch14](https://callous-0923.github.io/agent-study/chapter_14_sqlite/14_sqlite_agent_storage.html)** | SQLite Persistence | 5-table schema, WAL mode, session/task/user management |
| **[Ch15](https://callous-0923.github.io/agent-study/chapter_15_a2a/15_a2a_protocol.html)** | Google A2A Protocol | AgentCard, Task, Artifact, multi-agent collaboration |
| **[Ch16](https://callous-0923.github.io/agent-study/chapter_16_memgpt/16_memgpt_letta.html)** | MemGPT/Letta Memory | Core Memory, Heartbeat, Sleep-Time, Filesystem Memory |
| **[Ch17](https://callous-0923.github.io/agent-study/chapter_17_computer_use/17_computer_use.html)** | Computer Use | Screenshot-Action Loop, coordinate calculation, security sandbox |
| **[Ch18](https://callous-0923.github.io/agent-study/chapter_18_security/18_agent_security.html)** | Agent Security & Guardrails | Prompt Injection, privilege levels, input sanitization, 4-layer defense |

### Tier 5: Advanced Architecture & Optimization

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch19](https://callous-0923.github.io/agent-study/chapter_19_workflow_patterns/19_workflow_patterns.html)** | Agentic Workflow Patterns | Reflection / Routing / Orchestrator-Worker / Evaluator-Optimizer (7 patterns) |
| **[Ch20](https://callous-0923.github.io/agent-study/chapter_20_context_engineering/20_context_engineering.html)** | Context Engineering | Context Rot, budget management, XML-structured prompts, Skill.md |
| **[Ch21](https://callous-0923.github.io/agent-study/chapter_21_streaming/21_streaming_architecture.html)** | Streaming & Real-time Architecture | EventBus, dynamic interrupts, backpressure, StateManager Reducer |
| **[Ch22](https://callous-0923.github.io/agent-study/chapter_22_dspy/22_dspy.html)** | DSPy Auto-Optimization | Signature→Module→Optimizer, automatic few-shot, LangChain complement |
| **[Ch23](https://callous-0923.github.io/agent-study/chapter_23_code_agents/23_code_agents.html)** | Code Agent Architecture Survey | CodeAct / ACI / Plan-Execute, SWE-bench, Agentless findings |
| **[Ch24](https://callous-0923.github.io/agent-study/chapter_24_observability/24_observability.html)** | Agent Observability | Tracing Span tree, dashboards, LangSmith vs LangFuse, alerting |

### Tier 6: Foundational Reinforcement

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch25](https://callous-0923.github.io/agent-study/chapter_25_vectordb/25_vectordb.html)** | Vector Database Selection | Chroma / Pinecone / Milvus / Qdrant comparison, Embedding trade-offs |
| **[Ch26](https://callous-0923.github.io/agent-study/chapter_26_model_routing/26_model_routing.html)** | Model Routing Strategies | Threshold / Cascade / Semantic / Cost-Aware (4 strategies) |
| **[Ch27](https://callous-0923.github.io/agent-study/chapter_27_prompt_eng/27_prompt_engineering.html)** | Agent Prompt Engineering | System Prompt 6-module template, tool description scoring card |
| **[Ch28](https://callous-0923.github.io/agent-study/chapter_28_cache/28_cache.html)** | Semantic Caching & Token Optimization | Exact→Semantic→LLM three-tier cache, token budget management |

### Tier 7: Expert Level

| Ch | Topic | Key Technologies |
|------|------|----------|
| **[Ch29](https://callous-0923.github.io/agent-study/chapter_29_multimodal/29_multimodal.html)** | Multi-Modal Agent | Vision + text joint reasoning, multi-modal tool calling |
| **[Ch30](https://callous-0923.github.io/agent-study/chapter_30_reliability/30_reliability.html)** | Agent Reliability Engineering | Circuit breaker, exponential backoff retry, idempotency, degradation |
| **[Ch31](https://callous-0923.github.io/agent-study/chapter_31_benchmarks/31_benchmarks.html)** | Agent Benchmarking in Depth | GAIA / AgentBench / WebArena / tau-bench |
| **[Ch32](https://callous-0923.github.io/agent-study/chapter_32_self_improving/32_self_improving.html)** | Self-Improving Agent | Bad case collection → auto prompt optimization → eval validation |
| **[Ch33](https://callous-0923.github.io/agent-study/chapter_33_prompt_cache/33_prompt_cache.html)** | Prompt Caching & Inference Optimization | Anthropic Cache, KV sharing, speculative decoding |
| **[Ch34](https://callous-0923.github.io/agent-study/chapter_34_finetune/34_finetune.html)** | Model Fine-tuning for Function Calling | LoRA, fine-tuning data preparation, cost-benefit analysis |
| **[Ch35](https://callous-0923.github.io/agent-study/chapter_35_data_flywheel/35_data_flywheel.html)** | Data Flywheel | Interaction collection → bad case detection → auto-triggered improvement |
| **[Ch36](https://callous-0923.github.io/agent-study/chapter_36_defense/36_defense.html)** | Agent Defense in Depth | Canary Token, layered isolation, behavioral sandbox |

---

## 🚀 Quick Start

### 1. Clone the Repository

```bash
git clone https://github.com/Callous-0923/agent-study.git
cd agent-study
```

### 2. Check Environment

```bash
python chapter_00_overview/00_course_overview.py
```

### 3. Install Dependencies

Open `chapter_00_overview/00_course_overview.py`, change `install = False` to `install = True`, then run:

```bash
python chapter_00_overview/00_course_overview.py
```

### 4. Configure API Key (only needed for Ch1-5)

Create a `.env` file in the project root:

```env
OPENAI_API_KEY=sk-your-api-key-here
OPENAI_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-4o-mini
```

> You can also use other providers (DeepSeek, Qwen, etc.) by changing `OPENAI_BASE_URL` and `LLM_MODEL`.

### 5. Start Learning (most chapters require no API key!)

```bash
# No API key needed — run directly
python chapter_08_claude_code/08_claude_code_architecture.py
python chapter_10_mcp/10_mcp_deepdive.py
python chapter_14_sqlite/14_sqlite_agent_storage.py
python chapter_19_workflow_patterns/19_workflow_patterns.py
python chapter_21_streaming/21_streaming_architecture.py
python chapter_24_observability/24_observability.py
python chapter_26_model_routing/26_model_routing.py
python chapter_28_cache/28_cache.py
```

---

## 📦 Dependencies

| Dependency | Chapters | Description |
|------|----------|------|
| `openai` | Ch1-3 | OpenAI API calls |
| `langchain` + `langgraph` | Ch4-5 | Agent framework hands-on |
| `fastapi` + `uvicorn` | Ch13 | Agent service deployment |
| `pydantic` | Ch13 | Data model validation |
| `python-dotenv` | Ch0 | Environment variable management |

> **Note**: Chapters 8-12 and 14-28 mostly rely on the Python standard library (`sqlite3`, `asyncio`, `hashlib`, etc.) — no extra installation needed.

---

## 🎯 Learning Paths

### Path 1: From Scratch (recommended for beginners)
Follow Ch0 → Ch36 in order, 1-2 hours per chapter.

### Path 2: Interview Cram (key chapters)
- **Ch7**: 20 high-frequency interview questions + process
- **Ch8**: Claude Code architecture (industry-grade Agent design)
- **Ch9**: RAG deep dive
- **Ch10 + Ch15**: MCP & A2A dual protocols
- **Ch11**: Tool Calling internals
- **Ch18**: Agent security (demo vs. production engineer differentiator)
- **Ch19**: Workflow design patterns (universal system design framework)
- **Ch26**: Model routing (50-80% cost reduction)

### Path 3: Build Products
- **Ch13**: FastAPI service
- **Ch14**: SQLite persistence
- **Ch12 + Ch24**: Production checklist + observability
- **Ch26 + Ch28**: Cost optimization (routing + caching)

---

## 🧠 Core Technology Coverage

```
Tool Calling Internals  ★★★★★  Complete OpenAI vs Anthropic comparison + Streaming assembly
MCP Protocol            ★★★★★  Full lifecycle simulation (Initialize→tools/call)
A2A Protocol            ★★★★★  AgentCard/Task/Artifact + Multi-Agent collaboration
Claude Code Architecture★★★★★  nO/h2A/Compaction/SubAgent reverse analysis
RAG Full-Stack          ★★★★   Chunking/Embedding/RRF/Cross-Encoder/Production
Model Routing           ★★★★   4 strategies + cost comparison experiments (94% savings)
Semantic Caching        ★★★★   Three-tier cache + Token budget management
Agent Security          ★★★★   Prompt Injection + privilege levels + 4-layer defense
DSPy Auto-Optimization  ★★★★   Signature/Module/Optimizer + LangChain complement
Agentic Workflow        ★★★★   7 design patterns + system design answer framework
Context Engineering     ★★★★   Context Rot principles + XML Prompt + budget management
Streaming Architecture  ★★★★   EventBus + dynamic interrupts + backpressure
Observability           ★★★★   Tracing Span tree + LangSmith vs LangFuse
MemGPT Memory           ★★★★   Core Memory/Heartbeat/Sleep-Time/Filesystem
Code Agents             ★★★    CodeAct/ACI/Plan-Execute + SWE-bench
Vector Databases        ★★★    Chroma/Pinecone/Milvus/Qdrant + Embedding strategies
FastAPI Service         ★★★    REST/SSE/WebSocket + production deployment
SQLite Persistence      ★★★    5-table Schema + WAL + audit queries
Computer Use            ★★★    Screenshot-Action Loop + security sandbox
```

---

## 📁 Project Structure

```
agent-study/
├── README.md
├── README_EN.md
├── .gitignore
├── chapter_00_overview/              🚀 Course overview + setup
├── chapter_01_fundamentals/          📖 First Agent (handwritten ReAct)
├── chapter_02_components/            🧩 Planner + Memory + Tool design
├── chapter_03_types/                 🎯 ReAct / Plan-Execute / Reflexion
├── chapter_04_frameworks/            🔧 LangChain + LangGraph
├── chapter_05_multi_agent/           🤝 Multi-agent collaboration
├── chapter_06_evaluation/            📊 Evaluation + testing strategies
├── chapter_07_interview/             🎓 20 interview questions + career guide
├── chapter_08_claude_code/           🏗️ Claude Code architecture reverse engineering
├── chapter_09_rag_deepdive/          🔍 RAG deep dive (Chunking, Embedding, Production)
├── chapter_10_mcp/                   🔌 MCP protocol complete implementation
├── chapter_11_tool_calling/          ⚙️ Tool Calling internals
├── chapter_12_infrastructure/        🏭 OpenClaw/Harness/Production infrastructure
├── chapter_13_fastapi/               🌐 FastAPI Agent service
├── chapter_14_sqlite/                💾 SQLite persistence
├── chapter_15_a2a/                   🤖 Google A2A protocol
├── chapter_16_memgpt/                🧠 MemGPT/Letta memory architecture
├── chapter_17_computer_use/          🖥️ Computer Use + GUI
├── chapter_18_security/              🛡️ Agent security & guardrails
├── chapter_19_workflow_patterns/     🏷️ Agentic Workflow design patterns
├── chapter_20_context_engineering/   📐 Context Engineering
├── chapter_21_streaming/             📡 EventBus real-time architecture
├── chapter_22_dspy/                  🔬 DSPy auto-optimization
├── chapter_23_code_agents/           📊 Code Agent architecture survey
├── chapter_24_observability/         📈 Observability (LangSmith/LangFuse)
├── chapter_25_vectordb/              🗄️ Vector database selection
├── chapter_26_model_routing/         🔀 Model routing & cost optimization
├── chapter_27_prompt_eng/            ✍️ Agent Prompt engineering
├── chapter_28_cache/                 ⚡ Semantic caching & token optimization
├── chapter_29_multimodal/            👁️ Multi-Modal Agent
├── chapter_30_reliability/           🛡️ Agent reliability engineering
├── chapter_31_benchmarks/            📊 Agent benchmarking in depth
├── chapter_32_self_improving/        🔄 Self-Improving Agent
├── chapter_33_prompt_cache/          💾 Prompt Caching & inference optimization
├── chapter_34_finetune/              🎯 Model fine-tuning for Function Calling
├── chapter_35_data_flywheel/         🔁 Data flywheel
└── chapter_36_defense/               🏰 Agent defense in depth
```

---

## ✨ Features

- **📝 Lecture-as-Code**: Each `.py` file is both a complete lecture (module-level docstring) and runnable code
- **🤖 No API Key Needed**: Chapters 8-28 mostly use only the standard library — run directly
- **🎤 Interview-Oriented**: Each chapter highlights key interview topics with answer frameworks and scoring points
- **🔗 Cross-Referenced**: Chapters form a complete knowledge network through cross-references
- **📊 Runnable Demos**: Every chapter includes complete demonstration output
- **🇨🇳 Chinese-first** with English README available

---

## 🌟 References

- [ReAct: Synergizing Reasoning and Acting in Language Models](https://arxiv.org/abs/2210.03629) (Yao et al., ICLR 2023)
- [MemGPT: Towards LLMs as Operating Systems](https://arxiv.org/abs/2310.08560) (Packer et al., 2023)
- [Reflexion: Language Agents with Verbal Reinforcement Learning](https://arxiv.org/abs/2303.11366) (Shinn et al., 2023)
- [DSPy: Compiling Declarative LM Calls into Self-Improving Pipelines](https://arxiv.org/abs/2310.03714) (Khattab et al., 2023)
- [SWE-bench: Can Language Models Resolve Real-World GitHub Issues?](https://arxiv.org/abs/2310.06770) (Jimenez et al., 2024)
- [Agentless: Demystifying LLM-Based Software Engineering Agents](https://arxiv.org/abs/2407.01489) (Xia et al., 2024)
- [MCP Official Specification](https://modelcontextprotocol.io) (Anthropic, 2024-2025)
- [A2A Protocol Specification](https://a2a-protocol.org) (Google, 2025)
- [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) (Anthropic, 2024)
- [Effective context engineering for AI agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) (Anthropic, 2025)
- [Claude Code Reverse Analysis](https://github.com/shareAI-lab/analysis_claude_code) (Community, 2025)

---

## 📄 License

MIT License — Free to use, modify, and distribute.

---

<p align="center">
  <b>If this project helps you, please give it a ⭐ Star!</b><br>
  <sub>36 Chapters · 7 Tiers · Continuously updated · Issues and PRs welcome</sub>
</p>