smolagents

smolagents is a minimal library that lets you build powerful agents in a few lines of code. Its design goals:

• Simplicity. The core agent logic is ~1,000 lines — you can read all of it.
• Code agents first. CodeAgent writes its actions as Python code instead of JSON tool calls. Research shows this is more reliable and more token-efficient than JSON blobs.
• Model-agnostic. Use any LLM — an Inferix endpoint, a local transformers model, an OpenAI-compatible API, or 100+ providers via LiteLLM.
• Tool-agnostic. Wrap any Python function as a tool, or import tools from the Hub, LangChain, or an MCP server.
• Secure. Run agent-generated code in a sandbox (E2B, Docker, or WebAssembly).

How a code agent works

A CodeAgent runs a ReAct loop. At each step it writes a snippet of Python, that code is executed, and the observations feed the next step — until it calls final_answer(...).

Thought:  I should search for the current population, then compute the ratio.
Code:     pop = web_search("Guangzhou population 2024")
          print(pop)
Observation: 18.7 million ...

Continue to Installation, then the Quickstart.