Capability · Framework — agents

smolagents

smolagents is Hugging Face's compact agent library. Its main idea is CodeAgents — agents that reason by writing Python code snippets that call tools, rather than emitting JSON function calls. It's deliberately tiny (~1000 LOC), supports any LLM through Hugging Face's inference API or LiteLLM, and runs tools safely in a sandboxed Python interpreter or via E2B.

Framework facts

Category: agents
Language: Python
License: Apache 2.0
Repository: https://github.com/huggingface/smolagents

Install

pip install smolagents

Quickstart

from smolagents import CodeAgent, InferenceClientModel, DuckDuckGoSearchTool

model = InferenceClientModel()
agent = CodeAgent(tools=[DuckDuckGoSearchTool()], model=model)
agent.run('How many seconds in a leap year?')

Alternatives

LangGraph — lower-level graph framework
Agno — Python performance-focused alternative
CrewAI — role-based alternative
Pydantic AI — typed alternative

Frequently asked questions

Why write Python instead of JSON tool calls?

Code-as-action agents can compose multi-step logic in one pass — loops, conditionals, variable reuse — which JSON function-call agents can only express as separate back-and-forth turns. The smolagents/HF paper reports fewer steps and higher task success on many benchmarks.

Is it safe to run generated Python?

smolagents ships a local sandboxed Python interpreter that restricts imports and dangerous builtins, and it integrates with E2B for remote sandboxing. Still — treat it like any code execution: don't expose production secrets to untrusted prompts.

Sources

smolagents — docs — accessed 2026-04-20
smolagents — GitHub — accessed 2026-04-20