AI Agents: From Fundamentals to Certification (Hugging Face AI Agents Course)

Topics covered

• •Understand the definition of an AI agent and the spectrum of "agency" (from a simple processor to multi-agent systems).
• •Understand how LLMs work: tokens, special tokens, chat templates.
• •Master the agent's Thought → Action → Observation loop and the ReAct approach.

Topics covered

• •Understand the philosophy of smolagents: minimalism, a code-first approach, integration with the Hugging Face Hub.
• •Master the two agent types: `CodeAgent` (actions are Python code) and `ToolCallingAgent` (actions are JSON).
• •Learn to create tools via the `@tool` decorator and the `Tool` subclass, and to connect tools from the Hub, Spaces, LangChain, and MCP servers.

Topics covered

• •Understand how LlamaIndex is organized: components, tools, agents, and workflows; the role of LlamaHub.
• •Master the full RAG pipeline: loading documents → indexing → storing → querying → evaluation.
• •Learn to create four types of tools: FunctionTool, QueryEngineTool, Toolspecs, Utility Tools.

Topics covered

• •Understand when LangGraph is needed: the "control vs. freedom" trade-off and how it differs from LangChain.
• •Master the building blocks: State (TypedDict), Nodes, Edges (direct and conditional), StateGraph, START/END.
• •Build a first graph — an email classifier with conditional routing.

Topics covered

• •Understand how agentic RAG differs from classic RAG: the agent itself chooses the tool and answering strategy.
• •Build a BM25-based retriever tool over a dataset of guests.
• •Create helper tools: web search, weather (a dummy API), Hugging Face Hub statistics.

Topics covered

• •Understand what the GAIA benchmark is: its design principles, difficulty levels, and the "human vs. AI" gap.
• •Understand the final assignment's format: 20 Level 1 questions, the submission API, the leaderboard, the 30% threshold.
• •Design and build your own agent for solving GAIA tasks.

Topics covered

• •Understand function calling as a learned model capability, and fine-tune a model with LoRA (bonus 1).
• •Master agent observability and evaluation: traces, metrics, offline/online evaluation, LLM-as-a-Judge (bonus 2).
• •Understand how an agent differs from an "LLM in a game," and build an agent for Pokémon battles (bonus 3).