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

Era	Approach	Example
2024-2025	Chunk docs, embed, vector search, wire the agent loop yourself	rag-agent-demo/
2026 (SDKs)	Built-in tools + custom MCP servers, no glue code	claude-agent-sdk-demo/
2026 (Frameworks)	Type-safe agents with progressive skill disclosure	pydantic-ai-skills-demo/

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
• Multi-provider - Works with OpenRouter, OpenAI, or Ollama

.claude/skills/pptx-generator/SKILL.md

A real-world skill file showing how skills are structured as packaged expertise in markdown. This is the PowerPoint generator skill from the Dynamous Second Brain - a concrete example of the SKILL.md format discussed in the video.

diagram.excalidraw

The Excalidraw diagram used throughout the video covering:

1. The Old Way (2024-2025) - Pick framework, define tools, set up RAG, wire agent loop
2. Batteries-Included SDKs (2026) - Claude Agent SDK, Codex, built-in tools, skills vs tools
3. Frameworks Still Matter - Pydantic AI, LangGraph, OpenAI Agents, CrewAI
4. The Decision - When to use an SDK vs a framework
5. What Happened to RAG? - From naive RAG to agentic RAG and hybrid approaches

Links

• Claude Agent SDK
• Agent Skills Standard
• Pydantic AI
• LangGraph
• Anthropic Building Effective Agents
• LlamaIndex: Did Filesystem Tools Kill Vector Search?