Last week, Hugging Face unveiled a new coding library aimed at simplifying the development of artificial intelligence (AI) agents. Named smolagents, this tool outlines fundamental logic for creating straightforward AI agents capable of executing actions through code. Developers can integrate smolagents with various open-source large language models (LLMs) or select cloud-based LLMs, in addition to constructing tools to facilitate external output from the agents. These tools can also be shared with the community, allowing other developers to access and utilize them.
Hugging Face Launches Smolagents Library for AI Agents
In a recent blog post, the prominent AI and machine learning (ML) platform disclosed the new library’s features, which are designed to enhance the agentic capabilities for developers. The library is comprised of approximately 1,000 lines of code that establish the core functionalities of an AI agent. By enabling developers to connect it to an LLM and relevant tools for gathering external data or performing actions, Hugging Face asserts that this streamlined focus will ease the process of creating new agents for various projects and applications.
Smolagents are particularly tailored for simple tasks, meaning they are versatile but may not suit complex, multi-step, or multi-agent operations effectively. According to Hugging Face, while the system can execute coded actions, it is not designed to generate code itself. The platform further facilitates developers by allowing the execution of the AI agent in sandboxed environments through E2B, enabling reliability testing and output adjustments.
The library also includes support for the standard ToolCallingAgent, which drafts actions in JSON format or textual representations. Once developers create tools for the agent, they are encouraged to share these with the broader community. Users can choose from any open model available on the platform through a free inference application programming interface (API) or select from a vast array of over 100 cloud-based models.
Regarding the development process, Hugging Face advises that developers incorporate functions with type hints as well as detailed descriptions for inputs and outputs. To illustrate the library’s capabilities, the platform presented a code example for an AI agent designed to retrieve travel times from Google Maps and assist users in planning their itineraries.
