smolagents

Hugging Face designed smolagents around one idea: let the model express each step as a short Python snippet instead of a JSON tool call. Its CodeAgent runs a reason-act loop in which generated code calls your tools as ordinary functions, so loops, variables, and several tool calls fit into a single step, and a ToolCallingAgent covers providers that prefer structured calls. All agent logic lives in one readable module, yet multi-agent hierarchies, image inputs, and tools pulled from MCP servers, LangChain, or Hub Spaces are supported. Model classes cover local Transformers, vLLM, and MLX, Hugging Face Inference Providers, OpenAI-compatible servers such as Ollama, Azure OpenAI, Amazon Bedrock, and LiteLLM. Since the agent runs code it wrote, the maintainers stress that the bundled local Python executor is not a real security boundary and point users to Docker, E2B, Modal, or Blaxel sandboxes. Install from PyPI on Python 3.10 or newer; the smolagent and webagent commands provide a CLI and a browser-driving agent. Apache 2.0 permits commercial use. A good fit for Python developers who want to read and modify the whole agent loop.