Hugging Face releases smolagents: CodeAgent paradigm allows Agent to call tools in the form of code, supporting MCP and sandbox isolation
Hugging Face launched the lightweight Agent library smolagents, which takes the CodeAgent paradigm as the core and natively supports MCP tools, E2B sandbox isolation and HF Hub model sharing, emphasizing minimalist abstraction and fast access.
smolagents releases the latest capability updates to enhance enterprise implementation efficiency
Hugging Face launches smolagents, an open source Python Agent library with a minimalist design and a core code of about 1,000 lines. The biggest feature is that CodeAgent is used as a first-class citizen paradigm: Agent is not called through JSON description tools, but directly writes Python code to organize the control flow, and naturally supports function nesting, loops and conditional logic.
CodeAgent vs ToolCallingAgent
CodeAgent (recommended by default): Agent writes operations as Python code for execution, and tool calls are implemented through code combination. Suitable for complex logical task links.
ToolCallingAgent: Supports the traditional JSON/text tool calling paradigm and is compatible with mainstream tool description formats. Suitable for tool interface standardization scenarios.
Safe execution: built-in sandbox isolation support
CodeAgent execution requires security isolation, and smolagents natively connects to multiple sandbox solutions:
- E2B — Firecracker microVM isolation sandbox
- Modal — Serverless computing platform
- Docker — local isolation
- Blaxel — Cloud Agent runtime
Model independent × Tool independent
Supported models: HF Inference API, OpenAI, Anthropic, LiteLLM, native Transformers / Ollama.
Supported tools: MCP Server (native support), LangChain tool, HF Hub Space tool, custom Python function.
Quick installation
pip install 'smolagents[toolkit]'from smolagents import CodeAgent, InferenceClientModel, DuckDuckGoSearchTool
model = InferenceClientModel()
agent = CodeAgent(tools=[DuckDuckGoSearchTool()], model=model)
result = agent.run("What are the major developments in AI Agent in the past week?")HF Hub ecological integration
Smolagents supports uploading Agents and tools to HF Hub in the form of Gradio Space, and can also directly load Agent configurations shared by the community from the Hub. It is currently a rare Agent framework with the capability of "community asset sharing".
Official documentation: https://huggingface.co/docs/smolagents | Installation: pip install smolagents
Reviews