smolagents
used in
0
stacks
Hugging Face's minimalist agent library where agents act by writing and running code.
A small, low-abstraction library whose core logic is roughly a thousand lines, centered on "code agents" that express actions as executable Python rather than JSON tool calls. It is model-agnostic via LiteLLM and the Hugging Face Hub, supporting local and hosted models, and runs agent code in sandboxes such as E2B, Modal, or Docker. Tools and agents can be shared and pulled from the Hub.
Strengths
- Minimal, readable codebase with few abstractions
- Code-agent approach can reduce steps and LLM calls
- Model-agnostic via LiteLLM and the Hub
- Sandboxed execution options (E2B, Modal, Docker)
Tradeoffs
- Minimalism means fewer built-in features for complex orchestration
- Executing model-written code requires careful sandboxing
- Younger, smaller ecosystem than larger frameworks