Introduction to Hugging Face smolagents

AI Agents with Hugging Face smolagents

Adel Nehme

VP of AI Curriculum, DataCamp

AI agent definition

From Chatbots to Agents

Chatbots:

Respond with natural language
Passive and reactive

Agents:

Can take actions (web search, read files, analyze data, etc.)
Actively reason toward a goal

Agents follow a cycle:

Example: Competitor Pricing Research

All from a single prompt!

What Is smolagents?

Lightweight Python framework
Developed by Hugging Face

Supports two types of agents:

ToolCallingAgent: Uses structured function calls
CodeAgent: Writes and runs Python code

How Function-Calling Works

Action 1: {"tool": "search_company", "company": "Competitor A"}
Action 2: {"tool": "get_pricing", "company": "Competitor A", "plan": "Basic"}
Action 3: {"tool": "get_pricing", "company": "Competitor A", "plan": "Pro"}
Action 4: {"tool": "search_company", "company": "Competitor B"}
Action 5: {"tool": "get_pricing", "company": "Competitor B", "plan": "Basic"}

How Code Agents Work

competitors = ["Competitor A", "Competitor B", "Competitor C"]
pricing_data = {}

for company in competitors:
    company_info = search_company(company)
    plans = extract_pricing_plans(company_info)
    pricing_data[company] = plans

most_affordable_option = min(pricing_data, 
                      key=lambda x: pricing_data[x]['basic_plan'])

The Code Agent Flow

Research shows ~20% higher success rate than function-calling methods.

¹ https://huggingface.co/papers/2402.01030

Let's practice!

AI Agents with Hugging Face smolagents