Summary: This repository hosts the open-source content and code for the book "Deep Dive into AI Agents: Design Principles and Engineering Practices." It provides a comprehensive guide to building AI agents, centered on the formula Agent = LLM + Context + Tools. Aimed at engineers and researchers, it bridges theory with practical, chapter-aligned code examples for hands-on mastery.

How It Works: The project's core is the Agent = LLM + Context + Tools framework. The LLM serves as the reasoning engine, Context manages state and instructions, and Tools enable external interaction. The repository details advanced techniques across ten chapters, including context management (RAG, memory), tool integration (perception, execution, collaboration via MCP protocol), multi-modal capabilities, and multi-agent systems, emphasizing engineering practices beyond the base LLM.

Quick Start & Requirements:

• Setup: Compile the book PDF using pandoc, xelatex, and ElegantBook in book/ (bash build_pdf.sh). Code examples are per chapter (chapterN/project_name/).
• Prerequisites: Standard Python. API keys for LLM providers (Kimi, OpenRouter, etc.) recommended. GPU resources may be needed for training. Docker deployment options available for Chapter 4 tools.
• Resources: Links to API key applications provided.

Highlighted Details:

• Agent Framework: Deep dive into Agent = LLM + Context + Tools, covering fundamentals to advanced topics.
• MCP Protocol: Standardized approach for Perception, Execution, and Collaboration tools.
• Code Examples: Over 100 runnable projects demonstrating concepts from basic agents to multi-agent systems.
• Evaluation: Extensive coverage of agent evaluation benchmarks (Terminal-Bench, SWE-bench, GAIA, OSWorld).
• Self-Evolution: Mechanisms for agents to learn and improve without weight updates.

Maintenance & Community: Authored by Li Bojie. Contributions are welcomed via Pull Requests. No explicit community channels (Discord/Slack) or roadmap are detailed in the README.

Licensing & Compatibility: The code is for "learning reference only." Specific licenses vary by sub-project, requiring individual review for use beyond personal study. Commercial use or integration into closed-source projects is not explicitly addressed and may be restricted.

Limitations & Caveats: The primary caveat is licensing: code is for learning reference, and varied sub-project licenses may restrict broader use. PDF compilation requires specific LaTeX/pandoc installations. Running many examples necessitates obtaining and managing API keys for external LLM services.