smolagents

SmolAgents - Hugging Face AI Agent Framework

SmolAgents is a minimalist Python library (~1000 lines) for building AI agents that write and execute code. It emphasizes simplicity, security, and model-agnostic design.

Installation

pip install smolagents                    # Core library
pip install 'smolagents[toolkit]'         # With default tools (web search, etc.)
pip install 'smolagents[litellm]'         # LiteLLM for OpenAI/Anthropic
pip install 'smolagents[transformers]'    # Local model support
pip install 'smolagents[mcp]'             # MCP server integration
pip install 'smolagents[telemetry]'       # OpenTelemetry support

Core Concepts

Agent Types

CodeAgent - Primary agent that writes Python code to execute actions:

from smolagents import CodeAgent, InferenceClientModel

model = InferenceClientModel()  # Default: Qwen/Qwen2.5-Coder-32B-Instruct
agent = CodeAgent(tools=[], model=model)
result = agent.run("Calculate the first 20 Fibonacci numbers")

ToolCallingAgent - Uses JSON-based tool calls (no code execution):

from smolagents import ToolCallingAgent, InferenceClientModel

agent = ToolCallingAgent(tools=[], model=InferenceClientModel())
agent.run("Get the title of https://huggingface.co/blog")

Model Configuration

Hugging Face Inference API (recommended for quick start):

from smolagents import InferenceClientModel

# Default model
model = InferenceClientModel()

# Specific model with provider
model = InferenceClientModel(
    model_id="meta-llama/Llama-3.3-70B-Instruct",
    provider="together",  # or "sambanova", "fireworks", etc.
    token="YOUR_HF_TOKEN"
)

LiteLLM (100+ providers):

from smolagents import LiteLLMModel

# Anthropic
model = LiteLLMModel(
    model_id="anthropic/claude-3-5-sonnet-latest",
    api_key="YOUR_ANTHROPIC_API_KEY"
)

# OpenAI
model = LiteLLMModel(model_id="gpt-4o", api_key="YOUR_OPENAI_API_KEY")

Local Transformers:

from smolagents import TransformersModel

model = TransformersModel(
    model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
    max_new_tokens=4096,
    device_map="auto"
)