smolagents Review 2026

Hugging Face's minimalist agent framework — build code-writing agents in a few hundred lines of Python, with first-class support for open models.

smolagents is Hugging Face's small, opinionated agent framework designed to be the antidote to sprawling frameworks like LangChain. The entire library is around a thousand lines of Python you can actually read in one sitting. The core idea is 'code-writing agents': instead of the standard tool-calling loop where an LLM emits JSON that a runtime dispatches, smolagents has the model write and execute Python code that calls its tools directly. In practice this is dramatically more expressive — an agent can loop, branch, do arithmetic, chain tools in one step — and it maps well to what modern LLMs are actually best at (coding). The framework offers two execution paths: a locally sandboxed Python interpreter and E2B or Docker-backed remote sandboxes for untrusted code. Model support is broad: OpenAI, Anthropic, Google Gemini, any Hugging Face Inference or Transformers model, LiteLLM, Bedrock, and Ollama. Tools include a built-in web browser, DuckDuckGo/Google search, Python REPL, image generation, and any function decorated with @tool. smolagents is Model Context Protocol compatible on both sides — you can use an MCP server as a tool source, and you can expose a smolagents agent through an MCP endpoint. Because it favors open-weight models (via HF Transformers and Inference) as first-class citizens rather than an afterthought, smolagents is popular for research, education, and cost-sensitive production agents.