smolagents
Minimal agent library where agents write their actions as Python code.
Publisher documentation
Choose the model and code-execution boundary before running an agent
smolagents documents CodeAgent actions written as Python code, with local or hosted model integrations and optional sandbox execution. The README provides a toolkit installation extra and discusses Docker and hosted execution options; installing the library does not provide a model or a configured sandbox.
- smolagents repository: Apache-2.0. Library code terms do not include hosted inference, model weights, sandbox services or their separate licenses and charges.
The checked repository is Apache-2.0. Generated code and enabled tools can have consequential access; review executor, imports, filesystem/network permissions and model credentials before running. Local execution is not a verified security boundary. Hosted inference and sandbox services can incur separate charges; local models still need suitable compute. ASE did not execute generated code, test sandbox isolation or verify provider billing.
The review date is an ASE reminder, not an upstream guarantee.
Small code-writing agents you want to read end to end before trusting them.
You need a large integration ecosystem or managed orchestration out of the box.
About smolagents
Hugging Face's small agent library. Its CodeAgent writes actions as executable Python rather than JSON tool calls, with sandboxed execution options; a conventional tool-calling agent is also included. It is model-agnostic and deliberately small, favouring a readable core over a large feature surface.
smolagents is listed in this index, with 30k stars on GitHub. Its latest recorded repository push was in October 2026; that is not an ASE editorial review date.
