Last week, Hugging Face unveiled a new coding library designed to streamline the process of creating artificial intelligence (AI) agents. Known as smolagents, this tool establishes the foundational logic for general-purpose, simple AI agents that can execute tasks through code. Smolagents are compatible with any open-source large language model (LLM) as well as several cloud-based LLMs, allowing developers to design tools that connect to the external outputs of these agents. Moreover, these tools can be shared on the platform, making them accessible to other developers.
Hugging Face Introduces Smolagents Library for AI Agents
In a blog entry, the AI and machine learning (ML) platform highlighted that smolagents aims to facilitate the use of agentic capabilities for developers. The library consists of approximately 1,000 lines of code that outline the essential functionality of an AI agent. Developers can easily link this library with an LLM and append tools that gather external data or carry out actions. By concentrating on these two components, Hugging Face asserts that developers will find it more straightforward to create and implement new agents for their projects.
Smolagents are designed with simplicity in mind, enabling them to tackle various tasks; however, they may not be suitable for complex multi-step or multi-agent operations. While the platform clarified that the library can script and execute actions, it does not generate code itself. Additionally, developers can test the AI agents in sandboxed environments through E2B, which allows for reliability assessments and adjustments to the output.
This agent library also features a standard ToolCallingAgent that creates actions in JSON or text formats. Furthermore, once a developer crafts a tool for an agent, the sharing capabilities enable community access. Users have the option to select any open model hosted on the platform through a free inference application programming interface (API) or choose from over 100 available cloud-based models.
Regarding usage, Hugging Face recommends structuring functions with type hints for inputs and outputs, along with descriptions for those inputs. Highlighting a practical application, the platform demonstrated an example of an AI agent capable of retrieving travel times from Google Maps and assisting users in creating travel itineraries.
