Smolagents

Smolagents is a minimalist, lightweight AI agent library developed by Hugging Face. It enables building AI agents that think and act by generating and executing Python code snippets for tasks, allowing more flexible and composable workflows compared to traditional action representations like JSON or text commands.

The main features include code-based actions, where agents write executable Python code to perform tasks, enabling natural composition, state management, and complex logic. Simplicity is a key aspect, with the core library having a compact codebase designed for ease of understanding and debugging. It is model-agnostic, compatible with any large language model including those from Hugging Face Hub, OpenAI, Anthropic, and others via LiteLLM integration. Smolagents also offers secure execution by supporting code execution safely in sandboxed environments like E2B or Docker. It supports multi-agent and multi-step workflows, allowing chaining agents, multi-step reasoning, and parallel tool calls. Furthermore, it offers multiple input/output modalities, supporting text, images, video, and audio for diverse applications, and enables community sharing of tools and agents via the Hugging Face Hub.

You can use any model, including Hugging Face models like Qwen, OpenAI models, Anthropic, and others supported through LiteLLM and Transformers.

Smolagents supports two main types of agents: CodeAgent, which is the default style where the agent generates Python code snippets to perform actions, and ToolCallingAgent, which uses structured tool calls expressed in JSON or text for invoking external tools.

Smolagents includes a few essential tools to get started, such as DuckDuckGo Web Search, a Python Code Interpreter (built-in for CodeAgent), and a Transcriber (Speech-to-Text). Additionally, custom tools can be created with minimal effort to tailor agents for specific tasks.

Code generated by agents runs in secure sandboxed environments such as E2B or Docker containers. This isolation safeguards the host environment from malicious or erroneous code.

Smolagents excels at web searching and retrieval, data analysis via Python execution, automating complex workflows with multi-step reasoning, orchestrating multi-agent systems (such as planners delegating to specialist agents), and handling multi-modal inputs like images, videos, and audio.

To get started with Smolagents, you need to install it via pip using the command 'pip install smolagents huggingface_hub', then log in to Hugging Face with your API token. After that, define your agent by choosing an LLM and the necessary tools, and finally, run the agent with tasks such as web searches or code execution.