Agentic AI Practice

A collection of hands-on notebooks and projects exploring function calling, AI agents, multi-agent systems, and Model Context Protocol (MCP).

Contents

01 — Function Calling

Notebook	Description
01-Agentic_AI_OpenAI_Eng.ipynb	End-to-end agentic AI with OpenAI: function calling, parallel tool calls, agentic loops, structured outputs, Agents SDK, multi-agent handoffs, guardrails, shared context, and tracing
02-Function_Calling_Eng.ipynb	Function calling fundamentals — open-source (Mistral) vs proprietary (OpenAI) approaches
03-Function_Calling_Ukr_Lapa.ipynb	Function calling walkthrough (Ukrainian)

02 — Agents

Project	Description
anthropic_cu/	Computer-use agent using Anthropic's Claude — takes a task description and autonomously performs screen actions with step-by-step traces
ui-tars/	Computer-use agent using UI-TARS 1.5 (open-source, Qwen-based) — same autonomous screen interaction, self-hosted via HuggingFace
02-Multi-Agent-Systems.ipynb	Multi-agent system with a manager agent that delegates to specialized sub-agents (web search, summarization) using smolagents
03_Agentic_RAG.ipynb	Agentic RAG pipeline: planner → research → answer → judge agents over AG News and SQuAD knowledge bases with self-correction

03 — Model Context Protocol

Project	Description
MCP Setup Guide	Connect the Filesystem MCP server to Claude Desktop on macOS — enables Claude to read/write local files

Quick Start

git clone https://github.com/oleksa-kosovan/agent_practice.git
cd agent_practice
pip install -r requirements.txt

Set your API keys:

export OPENAI_API_KEY="your_key"
export ANTHROPIC_API_KEY="your_key"  # for computer-use agents

Open any notebook in Jupyter or Colab and run the cells.