smolagents

Why we included this project

smolagents keeps the agent loop small enough to read: the core logic fits in roughly a thousand lines of code, so you can trace what your agent does instead of trusting a black box. Its main idea is a CodeAgent that writes its next step as a Python snippet rather than emitting JSON tool calls, which makes multi-step reasoning easier to follow and debug. The library stays model-agnostic, working with local transformers or Ollama models, Hugging Face inference providers, or OpenAI and Anthropic endpoints through LiteLLM, and it can pull in tools from MCP servers, LangChain, or Hub Spaces. Code-running agents can be sandboxed with Docker, E2B, Modal, or Blaxel, and the CLI plus Hub sharing make it easy to prototype and pass work along. If you want a small, readable agent library you can extend yourself, this is a practical place to start.