Hugging Face Introduces Smolagents Library to Build AI Agents With Open-Source LLMs

Hugging Face introduced a new code library last week to enable developers to easily build artificial intelligence (AI) agents. Dubbed smolagents, the tool defines the basic logic for general purpose simple AI agents that can perform actions by executing them in code. Smolagent can be connected with any open-source large language model (LLM) or a select number of cloud-based LLMs. Developers can also build tools to connect the external output part of the agent. These tools can also be shared on the platform to let other developers access and use them.

Hugging Face Introduces Smolagents Library for AI Agents

In a blog postthe AI and machine learning (ML) platform announced the new tool that is aimed at making it easier for developers to use agentic capabilities. The library comes with roughly 1,000 lines of code that dictate the basic functionality of an AI agent. Developers can attach it with an LLM and any tools to collect external data or execute an action. By just focusing on these two elements, the platform claims that developers will find it easier to make new agents and use them in their projects and applications.

Smolagents is built with simple agents in mind. This means they can perform any task, but likely will not be a great fit for multi-step or multi-agent functions. Hugging Face stated that it can write actions in code (as in execute actions) but it cannot be used to write said code. The platform also allows developers to execute the AI agent in sandboxed environments via E2B to test reliability and tweak the output.

The agent library also supports standard ToolCallingAgent which writes actions in JSON or text blobs. Additionally, once a developer builds a tool for the agent, they can also share it with the community. Users can pick any open model hosted on the platform via a free inference application programming interface (API) or pick from a list of more than 100 different cloud-based models.

Coming to the tool, Hugging Face recommends making a function with type hints on inputs and outputs as well as descriptions for inputs. Highlighting a use case, the platform showcased code for an AI agent that can get travel times from Google Maps and plan travel itineraries for users.