Smolagents
github.com/huggingface/smolagents
Hugging Face's minimalist agent library where agents write and execute Python code as their primary reasoning tool, rather than calling predefined function schemas.
Why it is useful
Code-as-action is a more powerful and flexible tool use pattern than JSON function calling — the agent can compose arbitrary logic rather than being limited to predefined tools. The library is intentionally small (~1000 lines) which makes it auditable and easy to extend. Hugging Face's backing means it integrates cleanly with the model hub and transformers ecosystem.