三个工作实例表明AI剂开发是如何演变的 – – 从传统的基于RAG的方法发展到现代电池包括自毁K和技能框架。

□ 进化

| E Approach Approach 来来来去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去去 |-|-|-|-.

• 2024-2025 * * * * 春克文件,嵌入,向量搜索,给代理机自己接通 * * rag-agent-demo/ * * * * * ** ** 2026 (SDKs) ** + 内置工具 + 自定义 MCP 服务器, 无胶水代码 + + claude-agent-sdk-demo/] + ==
• ** 2026 (框架) ** 具有渐进技能披露的型号安全剂 * * pydantic-ai-skills-demo/

□ 实例

[[T3] – – 传统办法(蟒蛇)

经典的RAG模式是用于构建AI代理的默认游戏本. 显示完整的管道:块文档,生成嵌入,存储在矢量数据库中,检索相关的块,反馈到LLM.

** Stack:** Pydantic AI + PostgreSQL/pgvector + OpenAI 嵌入

有了大文件Corpora 仍然是正确的决定 但对于许多情况来说,下文中较新的办法完全不需要这种基础设施。

QQ claude-agent-sdk-demo/ - 电池- 包含的 SDK( TypeScript)

How Building AI Agents Has Completely Changed

Three working examples showing how AI agent development has evolved - from the traditional RAG-based approach to modern batteries-included SDKs and skill-based frameworks.

The Evolution

Examples

rag-agent-demo/ - The Traditional Approach (Python)

The classic RAG pattern that was the default playbook for building AI agents. Shows the full pipeline: chunk documents, generate embeddings, store in a vector database, retrieve relevant chunks, feed to an LLM.

cd rag-agent-demo
cp .env.example .env     # Add your OpenAI key + database URL
docker compose up -d     # Start PostgreSQL with pgvector
uv sync
uv run python -m ingestion.ingest --documents documents/ --clean
uv run python cli.py

Stack: Pydantic AI + PostgreSQL/pgvector + OpenAI embeddings

This works - and it’s still the right call when you have large document corpora. But for many use cases, the newer approaches below eliminate the need for this infrastructure entirely.

claude-agent-sdk-demo/ - Batteries-Included SDK (TypeScript)

A research agent that comes with tools out of the box. No RAG pipeline, no vector database, no embedding infrastructure. The SDK provides Read, Write, WebSearch, Bash, and Grep as built-in tools, and you can add your own via MCP servers.

cd claude-agent-sdk-demo
bun install
bun run agent.ts "How AI agent frameworks evolved in 2026"

What this demonstrates:

• Built-in tools - Read, Write, WebSearch, Bash, Grep with zero setup
• Custom MCP server - createSdkMcpServer() with save_note and search_notes tools
• Subagents - Researcher and Writer agents that the orchestrator delegates to
• Hooks - Real-time monitoring of tool usage

Uses your local Claude Code CLI credentials - no API key needed.

pydantic-ai-skills-demo/ - Framework with Skills (Python)

A Pydantic AI agent with a skill system implementing progressive disclosure. Instead of loading all instructions upfront, skills are discovered at ~100 tokens each and loaded on demand - letting an agent access hundreds of capabilities without overwhelming its context window.

cd pydantic-ai-skills-demo
cp .env.example .env     # Add your API key
uv sync
uv run python -m src.cli

What this demonstrates:

• Progressive disclosure - Skills load in 3 levels (metadata -> instructions -> resources)
• 5 working skills - Weather, research assistant, recipe finder, world clock, code review
• Type safety - Full Pydantic models and typed dependencies