The Hugging Face Agents Course is the open-source, certified path into agent engineering, built by the same Hugging Face team that maintains smolagents and a meaningful chunk of the open-source LLM tooling ecosystem. It’s free, it ends with a benchmarked final project, and it ships a real certificate if you pass. Hard to beat as an on-ramp if you’re already in the HF ecosystem.
View Hugging Face Agents Course on GitHub
- 30,116 stars
- 2,157 forks
- MDX
- Apache-2.0 License
Hugging Face Agents Course: Four units, ending in a real benchmark
The structure is more focused than the typical “intro to agents” curriculum. Unit 1 covers what agents actually are (LLMs, the model family tree, special tokens) without drowning you in theory. Unit 2 walks three frameworks side by side: smolagents (HF’s own, lightweight), LlamaIndex (data and indexes), and LangGraph (production graphs). Unit 3 is agentic RAG. Unit 4 is the payoff: a final project graded against an automated benchmark, with a public leaderboard of student results.
That final unit matters. Most agent courses hand you a certificate of completion; this one makes you ship something that’s scored against everyone else’s. It’s a different motivation.
smolagents gets first-class treatment
Hugging Face maintains smolagents, so it’s no surprise the course teaches it as the entry framework. What is a surprise is how honestly the other frameworks are covered: LangGraph and LlamaIndex each get a dedicated sub-unit (2.2 and 2.3) with the same depth, not a token paragraph. You leave the course able to make a real framework choice, not just fluent in HF’s house brand.
The bonus units are worth flagging too. Bonus Unit 1 covers fine-tuning an LLM for function calling, Bonus Unit 2 covers observability and evaluation (smolagents’ built-in tracing plus external tools), and Bonus Unit 3 is “Agents in Games with Pokemon”, a playful final deep-dive that’s surprisingly instructive for reinforcement-flavored agent design.
The HF ecosystem is the substrate
This is the honest trade-off. The course is built on HF infrastructure: the Hub for models and datasets, HF Spaces for hosting, smolagents for the default framework. If your day job already touches HF, that’s a feature: the lessons plug directly into your existing stack. If you live entirely in OpenAI/Anthropic APIs, expect to translate the code samples to your provider. The concepts transfer, the boilerplate doesn’t.
Prerequisites are modest: basic Python and basic LLM familiarity. If you don’t have the second yet, do Generative AI for Beginners first.
The certification actually means something
The final-unit benchmark isn’t a participation trophy. Your agent has to clear a real score on a held-out task set, and your result is published alongside other students’. The certificate you get at the end is a credible signal because it’s benchmarked, not just “watched all the videos”. For a free course, that’s rare.
It also means the course ends, which is its own kind of value. Open-ended curricula tend to get abandoned halfway; this one has a finish line and a verifiable reward on the other side.
Get started
Sign up free at bit.ly/hf-learn-agents, then start at Unit 0:
git clone https://github.com/huggingface/agents-course.gitThe repo is the source-of-truth mirror; the canonical reading experience is on the web at hf.co/learn/agents-course.
The Hugging Face Agents Course is the rare free, certified, benchmarked agent curriculum. If you want a structured path that ends in a credential hiring managers might actually recognize, this is the one: start at Unit 0, finish at Unit 4, ship the project.