Agent Development in Practice

Introduction

• Source: Andrew Ng Agentic AI Series、LangGraph Official Documentation、Anthropic Building Effective Agents
• Prerequisites: Python basics, understanding of LLM API calls, basic understanding of RAG concepts
• Programming language: Python
• Course difficulty: 🌟🌟🌟🌟
• Estimated study hours: 30+

AI Agents are the most central direction in AI application development today. If RAG solves "what the model knows," Agents solve "what the model does" — they can autonomously plan, use tools, and execute multi-step tasks, not just answer questions.

Core components of an Agent system: - LLM (brain): Responsible for reasoning and decision-making - Tools (hands): Search, code execution, API calls, file operations, etc. - Memory (notebook): Short-term (conversation history) + Long-term (persistent knowledge) - Planning (chain of thought): Decomposing complex tasks into executable steps

Andrew Ng repeatedly emphasized in his 2024-2025 course series: Agentic workflows are the most important paradigm shift in AI application development. Moving from simple single API calls to Agents that can autonomously complete multi-step tasks is a capability every AI engineer must master.

Course Resources

• Course: Andrew Ng - AI Agents in LangGraph — Free short course, hands-on with the LangGraph framework
• Documentation: LangGraph Documentation — The most mainstream multi-agent orchestration framework
• Article: Anthropic - Building Effective Agents — Anthropic's official Agent design guide, required reading

Resource Summary

Beginner (understanding core Agent concepts):

• Andrew Ng - Agentic Design Patterns (4 courses) — Covers Reflection, Tool Use, Planning, and Multi-Agent patterns, highly recommended
• Lilian Weng - LLM Powered Autonomous Agents — One of the most comprehensive Agent technical surveys
• ReAct: Synergizing Reasoning and Acting in LLMs — The original ReAct paper, understand the reasoning+action loop paradigm

Framework practice (pick one to go deep):

• LangGraph — State-graph-based Agent orchestration framework, supports loops, conditional branches, human-in-the-loop, the most flexible framework available
• CrewAI — Multi-agent collaboration framework with clear role division, quick to get started
• AutoGen (Microsoft) — Microsoft's open-source multi-agent conversation framework
• OpenAI Agents SDK — OpenAI's official Agent SDK, lightweight
• smolagents (HuggingFace) — Minimalist Agent framework, small codebase, great for learning principles

Tool Use (Tool Use / Function Calling):

• OpenAI Function Calling Guide — Understand how LLMs call external tools
• LangChain Tools & Toolkits — Rich collection of pre-built tools
• MCP (Model Context Protocol) — Anthropic's standardized tool protocol, the most important Agent infrastructure of 2025

Multi-Agent Systems:

• Andrew Ng - Multi-Agent Collaboration — CrewAI framework + multi-agent collaboration
• AutoGen Agent Chat — Multi-agent conversation and collaboration
• MetaGPT — Agent system simulating multi-role software company collaboration

Memory Systems:

• LangMem (LangChain Memory) — LangChain's official long-term memory solution
• Mem0 — Intelligent memory layer designed specifically for AI Agents
• Letta (MemGPT) — Brings OS memory management concepts into Agent memory

Notes

Agent development has a steep learning curve. Recommended path: first understand the ReAct paradigm (reasoning+action loop) → build a single Agent with LangGraph → add tool calls → add memory → try multi-agent collaboration.

Don't start with advanced frameworks. First hand-write a simple Agent with pure Python + OpenAI API (~100 lines of code) to understand what's happening at each step, then use frameworks to improve efficiency. This is the same principle as learning programming — write it by hand first, then use libraries.