smolagents

smolagents

smolagents is a barebones library from Hugging Face for building agents that think in code. The entire agent logic fits in ~1000 lines of code. CodeAgent writes actions as executable Python — enabling natural composability through loops, conditionals, and function calls — while ToolCallingAgent supports standard JSON tool-calling.

Key Features

• Code-first agents: Actions are Python snippets executed in a sandboxed interpreter, not JSON blobs
• Secure sandboxing: Execute code safely via Modal, E2B, Docker, or Blaxel
• Any model: Works with Hugging Face models, OpenAI, Anthropic, local Ollama, and any LiteLLM-compatible provider
• Hub integration: Share and load agents and tools directly as Hugging Face Spaces
• Multi-agent: Compose agents hierarchically with a manager orchestrating specialists
• Minimal surface area: Easy to audit, extend, and debug

Quick Start

pip install smolagents

from smolagents import CodeAgent, DuckDuckGoSearchTool, HfApiModel

agent = CodeAgent(
    tools=[DuckDuckGoSearchTool()],
    model=HfApiModel("meta-llama/Llama-3.3-70B-Instruct")
)
agent.run("What were the biggest AI breakthroughs of 2025?")

Add to ai-supply

npx ai-supply add smolagents-code-agents

Curated mirror of the open-source smolagents (Apache-2.0). Get it from the source.