Hugging Face AI Agents Course certificate

Learning plan

About 30 hours of focused study, split into 4 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 3 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Working Python (functions, classes, pip, virtual environments) and basic familiarity with LLM chat APIs
• A free Hugging Face account and access token. A free-tier inference provider or a local Ollama model is enough for most notebooks

1. Week 1 · 6 h

   Agent fundamentals and the fundamentals certificate (Unit 1)

   • Explain the Thought-Action-Observation loop and the difference between ReAct-style text actions and JSON or code actions
   • Show how chat templates, special tokens and system messages turn a conversation into the prompt the model actually sees
   • Define a tool (name, description, typed arguments, output) and explain how its schema reaches the model
   • Pass the Unit 1 quiz to earn the free fundamentals certificate

   Hands-on: Build the Unit 1 "First Agent" Space from the template, then add your own tool (for example a timezone converter or a unit-conversion function) and get the agent to call it correctly from a natural-language request.

   Use: Agents Course – Unit 1 Introduction

2. Weeks 1–2 · 10 h

   Frameworks: smolagents, LlamaIndex and LangGraph (Unit 2)

   • Build a CodeAgent and a ToolCallingAgent in smolagents and explain when code actions beat JSON tool calls
   • Use LlamaIndex components, tools and workflows to build a retrieval-backed agent
   • Model an agent as a LangGraph state graph with nodes, edges and conditional routing, and add a human-in-the-loop step
   • Compare the three frameworks on control, verbosity and observability

   Hands-on: Build the same "triage my inbox" agent twice, once in smolagents and once in LangGraph, using three stub tools (read_emails, label, draft_reply). Write a short comparison of the number of steps, tokens used and how easy each was to debug.

   Use: Unit 2.1 smolagents · Unit 2.2 LlamaIndex · Unit 2.3 LangGraph

3. Weeks 2–3 · 5 h

   Agentic RAG use case (Unit 3)

   • Build a retrieval tool over a custom dataset and let the agent decide when to retrieve, search the web or answer directly
   • Combine several tools (retriever, web search, weather or Hub stats) in one agent and handle tool errors

   Hands-on: Complete the Unit 3 gala-guest assistant, then swap in your own dataset (for example your team's docs) and test 10 questions. Log which ones the agent answers with the wrong tool.

   Use: Unit 3 Agentic RAG

4. Week 3 · 9 h

   Final assignment on the GAIA benchmark (Unit 4)

   • Build an agent with web search, file and code-execution tools that can answer GAIA level-1 style questions
   • Iterate using traces to find failure modes such as wrong tool choice, formatting errors and looping
   • Score 30% or higher (6 of 20) on the course's GAIA level-1 subset and submit to the Students leaderboard for the certificate of completion

   Use: Unit 4 Final Assignment · Agents Course Students leaderboard · Bonus Unit 2 – Agent observability and evaluation

You're ready to book when…

• You have the Unit 1 fundamentals certificate
• You can explain what your agent did at each step from its trace, not just its final answer
• Your agent scores 30% or more on the GAIA subset and appears on the leaderboard

Exam tips

• There's no deadline. The fundamentals certificate needs only Unit 1, while the full certificate needs Unit 1, a use-case assignment and the final challenge.
• The final test uses 20 level-1 GAIA validation questions, scored by exact match. Have your agent return only the answer, without the text "FINAL ANSWER", or correct answers will be marked wrong.