README.md

freeact

A lightweight library for code-action based agents.

Contents

• Introduction
• Key capabilities
• Quickstart
• Evaluation
• Supported models

Introduction

freeact is a lightweight agent library that empowers language models to act as autonomous agents through executable code actions. By enabling agents to express their actions directly in code rather than through constrained formats like JSON, freeact provides a flexible and powerful approach to solving complex, open-ended problems that require dynamic solution paths.

The library builds upon recent research demonstrating that code-based actions significantly outperform traditional agent approaches, with studies showing up to 20% higher success rates compared to conventional methods. While existing solutions often restrict agents to predefined tool sets, freeact removes these limitations by allowing agents to leverage the full power of the Python ecosystem, dynamically installing and utilizing any required libraries as needed.

Key capabilities

freeact agents can autonomously improve their actions through learning from environmental feedback, execution results, and human guidance. They can store and reuse successful code actions as custom skills in long-term memory. These skills can be composed and interactively refined to build increasingly sophisticated capabilities, enabling efficient scaling to complex tasks.

freeact executes all code actions within ipybox, a secure execution environment built on IPython and Docker that can also be deployed locally. This ensures safe execution of dynamically generated code while maintaining full access to the Python ecosystem. Combined with its lightweight and extensible architecture, freeact provides a robust foundation for building adaptable AI agents that can resolve real-world challenges requiring dynamic problem-solving approaches.

Quickstart

Install freeact using pip:

pip install freeact

Create a .env file with Anthropic and Gemini API keys:

# Required for Claude 3.5 Sonnet
ANTHROPIC_API_KEY=...

# Required for generative Google Search via Gemini 2
GOOGLE_API_KEY=...

Launch a freeact agent with generative Google Search skill using the CLI:

python -m freeact.cli \
  --model-name=claude-3-5-sonnet-20241022 \
  --ipybox-tag=ghcr.io/gradion-ai/ipybox:basic \
  --skill-modules=freeact_skills.search.google.stream.api

or an equivalent quickstart.py script:

import asyncio

from rich.console import Console

from freeact import Claude, CodeActAgent, execution_environment
from freeact.cli.utils import stream_conversation


async def main():
    async with execution_environment(
        ipybox_tag="ghcr.io/gradion-ai/ipybox:basic",
    ) as env:
        skill_sources = await env.executor.get_module_sources(
            module_names=["freeact_skills.search.google.stream.api"],
        )

        model = Claude(model_name="claude-3-5-sonnet-20241022", logger=env.logger)
        agent = CodeActAgent(model=model, executor=env.executor)
        await stream_conversation(agent, console=Console(), skill_sources=skill_sources)


if __name__ == "__main__":
    asyncio.run(main())

Once launched, you can start interacting with the agent:

freeact_iss_coffee_720.mp4

Evaluation

We evaluated freeact with the following models:

• Claude 3.5 Sonnet (claude-3-5-sonnet-20241022)
• Claude 3.5 Haiku (claude-3-5-haiku-20241022)
• Gemini 2.0 Flash (gemini-2.0-flash-exp)
• Qwen 2.5 Coder 32B Instruct (qwen2p5-coder-32b-instruct)
• DeepSeek V3 (deepseek-v3)
• DeepSeek R1 (deepseek-r1)

The evaluation uses two datasets:

1. m-ric/agents_medium_benchmark_2
2. m-ric/smol_agents_benchmark

Both datasets were created by the smolagents team at 🤗 Hugging Face and contain curated tasks from GAIA, GSM8K, SimpleQA, and MATH. We selected these datasets primarily for a quick evaluation of relative performance between models in a freeact setup, with the additional benefit of enabling comparisons with smolagents. To ensure fair comparisons with their published results, we used identical evaluation protocols and tools.

When comparing our results with smolagents using Claude 3.5 Sonnet on m-ric/agents_medium_benchmark_2 (only dataset with available smolagents reference data), we observed the following outcomes (evaluation conducted on 2025-01-07):

Interestingly, these results were achieved using zero-shot prompting in freeact, while the smolagents implementation utilizes few-shot prompting. You can find all evaluation details here.

Supported models

In addition to all supported models, freeact also supports the integration of new models from any provider that is compatible with the OpenAI Python SDK, including open models deployed locally with ollama or TGI, for example.