Welcome to "Building Code Agents with Hugging Face Smolagents", built in partnership with Hugging Face. This course shows you how to build code agents. Code agents are agents that write code to perform sequences of actions. This is distinct from coding agents, which are agents like those in Windsurf or Cursor that will write code for you to execute. I'm delighted that the instructors for this course are Thomas Wolf, who is co-founder and CSO of Hugging Face, who'll join you in the next lesson, as well as Aymeric Roucher, who is developer of smolagents at Hugging Face. Thanks, Andrew. I'm really excited about this course. I use LLMs to write code for me all the time to get stuff done. The idea behind this course is, let your LLM writes code for your agent as well to get stuff done. Code agents lets you take advantage of LLMs coding ability to write code to perform actions. So rather than having an LLM generate function calls one after another to complete a complex sequence of tasks where you generate one function, call it and execute that, then have the LLM decide what's the second function call to generate and execute that, and so on, one at a time. In contrast, a code agent will consolidate all of these calls into a single snippet of code. So lets an LLM lay out an entire plan of action all at the same time. That can then be executed efficiently rather than forcing an LLM to reveal the plan to you, one small step at a time. And it turns out to be more efficient and also give more reliable results. In the first lesson, Thomas will describe a brief history of agents to bring you to the current day. Then in the second lesson, you will learn about Hugging Face smolagents and get a chance to explore the benefits of code agents yourself. We will also review some of the academic results showing the savings of this approach. It is important when using your LLM-generated code to protect your system from potential ill effects. For example, the code may have syntax errors or may perform actions that could harm your system. To protect against this, in lesson three, you will learn about a constrained Python interpreter using smolagents and also how to execute your code in a sandbox for safety. In lesson four, you will see how to trace your agents in order to debug more complex scenarios. And in lesson five, you'll learn about and build a multi-agent system. Many people have worked to create this course. I'd like to thank from HuggingFace, Albert Villanova and from DeepLearning.AI, Geoff Ladwig also contributed to this course. Additionally, we receive help and support from Vasek Mlejnsky, co-founder and CEO of E2B, a company which provides cloud based secure sandboxes which you use today. All right, let's get started on the first lesson.
Building Code Agents with Hugging Face smolagents