Skip to content

agent-tutorial-README_EN.md

Latest commit

 

History

History
197 lines (153 loc) · 7.73 KB

agent-tutorial-README_EN.md

File metadata and controls

197 lines (153 loc) · 7.73 KB
Raw

AI Agent Zero-to-Hero Tutorial

What tech stack do agent products actually use? This guide breaks down the complete tech stack behind mainstream AI agents. Build your mental map from zero.

Overview: The Six-Layer Agent Tech Stack

+-----------------------------------------+
|  6. Observability  Tracing, debugging   |
+-----------------------------------------+
|  5. Deployment     Serve as API/UI      |
+-----------------------------------------+
|  4. Low-Code       Drag-and-drop        |
+-----------------------------------------+
|  3. Tools & Proto  Interact with world  |
+-----------------------------------------+
|  2. Memory         Short + long term    |
+-----------------------------------------+
|  1. Orchestration  Agent skeleton       |
+-----------------------------------------+
|  0. LLM            The brain            |
+-----------------------------------------+

0. LLM Providers -- The Brain

Updated 2026-07-22.

Global

Model Provider Positioning Key Strengths
GPT-5.5 OpenAI Flagship Apr 2026; 1M context; built-in Computer Use / MCP / Web Search; GPT-5.5 Pro for harder problems
Claude Opus 4.8 Anthropic Agent specialist May 2026; best for coding + long-running agentic tasks; Fable 5 (Jun) even stronger but access-restricted
Claude Sonnet 4.6 Anthropic Value pick Feb 2026; speed + intelligence balanced; 1M context (Beta); daily agent go-to
Gemini 3.6 Flash Google Low-cost agents Jul 22, 2026; 17% fewer tokens; optimized for large-scale agent workflows
Llama 4 Meta Open-source Local deployment; data stays on your machine

Note: GPT-4o, o3, o4-mini retired Feb 2026. Claude 3.5 also out of main lineup.

China

Model Provider Positioning Key Strengths
DeepSeek V4-Pro DeepSeek Open-source flagship Apr 2026; 1.6T params (MoE); 1M context; MIT license; Huawei Ascend native; ~1/10 GPT cost
Qwen 3.7-Max Alibaba CN #1 Arena blind-test leader among Chinese models; 100M+ downloads
GLM-5.2 Zhipu AI Long-horizon agents Jul 2026; million-token context; bi-monthly iteration; 6M+ global users
Kimi K2.6 Moonshot AI Agent Swarm Hundreds of sub-agents simultaneously; K2.5 does end-to-end software dev
MiniMax M3 MiniMax Cost-breaker Jun 2026; native multimodal; SWE-Bench Pro 59; ~1/5 Claude price
Wenxin 6.0 Baidu Massive scale 2.4T params; 2M context; offline on mobile
Hunyuan Hy3 Tencent Ecosystem 2.95B params (MoE); Apache 2.0; embedded in WeChat/Docs/Meetings
Baichuan M4 Baichuan Medical Medical-focused; AI family doctor; ultra-low hallucination

Note: DeepSeek R2 never shipped; reasoning absorbed into V4 Thinking mode.

How to choose?

Scenario Recommendation
Prototype / learn agent dev DeepSeek V4 (nearly free) or GPT-5.5
Coding agent Claude Opus 4.8 or GPT-5.3-Codex
High-volume, cost-sensitive Gemini 3.6 Flash or MiniMax M3
China-compliant production Qwen 3.7 / GLM-5.2
Long-horizon multi-step GLM-5.2 or Kimi K2.6
Air-gapped / data sovereignty DeepSeek V4 (MIT) or Llama 4
Healthcare / legal vertical Baichuan M4 or custom fine-tuned

1. Orchestration Frameworks -- The Skeleton

Framework Language Level One-liner Stars
LangChain Python/JS Entry Swiss Army knife; largest ecosystem 100K+
LangGraph Python/JS Mid State-machine agent flows; complex multi-step 10K+
LlamaIndex Python/JS Entry Data-first; indexing & RAG expert 38K+
CrewAI Python Entry Multi-agent teams with roles & goals 25K+
AutoGen Python Mid Microsoft; conversational multi-agent; enterprise 38K+
Semantic Kernel C#/Python/Java Mid Microsoft; deep Azure / .NET integration 22K+
Vercel AI SDK TypeScript Entry Frontend-first; streaming UX; Next.js 12K+
Spring AI Java Entry Java's LangChain; Spring Boot native 4K+
Haystack Python Entry NLP pipeline; search/QA agents 18K+
MetaGPT Python Mid Simulates software company SOP 50K+

2. Memory & Storage

Memory Type What it is Storage
Short-term Current conversation LLM context window
Long-term History, preferences Vector DB + SQL/NoSQL

Vector Databases

Product Type One-liner
Chroma Lightweight OSS pip install and go; learning & prototypes
Milvus Production OSS China's most mature; trillion-scale
Qdrant High-perf OSS Rust-based; excellent single-node
Weaviate Full-featured OSS Built-in vectorization + hybrid search
Pinecone Cloud Zero-ops; pay-per-use
pgvector PG extension Already on Postgres? One extension
Elasticsearch Search engine 8.0+ native vectors

3. Tools & Protocols -- Letting Agents Act

Mechanism Origin Description
Function Calling OpenAI Structured JSON output to call your functions
Tool Use Anthropic Define tools with name/description/schema
MCP Anthropic Industry standard -- "USB-C for agents"

4. Low-Code Platforms

Platform Provider Best for
Dify OSS / CN Visual workflows; built-in RAG; self-hosted
Coze (Kouzi) ByteDance Bot store; plugin market; WeChat/Douyin
FastGPT OSS / CN Knowledge-base Q&A; simple deploy
Alibaba Bailian Alibaba Cloud Qwen family; full pipeline
Baidu Qianfan Baidu Wenxin ecosystem
Flowise OSS Visual LangChain; drag-and-drop
n8n + AI OSS Automation + AI nodes

5. Deployment

Mode Tools Best for
API service LangServe, FastAPI REST API wrapper
Chat UI Chainlit, Gradio Quick demo
Serverless Cloudflare Workers, Vercel Zero-ops; lightweight
Containers Docker + K8s Production standard
Embedded Direct import Module in existing project

6. Observability

Tracing

Tool One-liner
LangSmith LangChain official; auto-records every call
LangFuse Open-source alternative; self-hostable
Phoenix (Arize) LLM observability; visual tracing
Weights & Biases MLOps veteran; LLM tracking

Evaluation

Dimension Method
Accuracy Human / LLM-as-judge
Tool-call correctness Audit right tool + right params
Latency End-to-end (LLM + tools)
Cost Token-usage tracking
Safety Jailbreak / injection detection

Scenario Quick-Reference

I want to build... Recommended stack
Personal AI assistant GPT-5.5 / DeepSeek V4 + LangChain + Tavily
Enterprise knowledge QA Dify or LlamaIndex + Milvus
Code review agent Claude Opus 4.8 + GitHub API + LangChain
Multi-role writing team CrewAI + Claude / DeepSeek
Data analyst agent LangChain + Python REPL + Streamlit
Customer service system LangGraph + PostgreSQL + MCP
Browser automation Playwright + Claude/GPT (built-in Computer Use)
AI chat app frontend Vercel AI SDK + Next.js
Java AI integration Spring AI + model APIs
High-volume agents Gemini 3.6 Flash + Dify / LangChain

Learning Path

Week 1: Call a model API
Week 2: First Function Calling with LangChain
Week 3: Add memory -- Chroma + RAG
Week 4: Learn MCP
Week 5: Deploy -- FastAPI or Dify
Week 6+: Multi-agent, fine-tuning, evaluation