Download
code-act is a research framework for building intelligent language-model agents that interact with their environment through executable code actions. The system proposes a unified action representation where language models produce Python code that can be executed directly, allowing the model to interact with external tools and environments in a structured way. By integrating a Python interpreter with the agent architecture, the system enables the agent to execute code, observe the results, and iteratively refine its actions through multiple reasoning steps. This approach helps unify reasoning and action planning within large language model agents by using code as the primary interface between the model and the external world. The framework also includes training data, models, and evaluation tools designed to study how language models can become more capable autonomous agents.
Features
- Executable code actions used as the unified action interface for LLM agents
- Python interpreter integration enabling real code execution and feedback
- Multi-turn reasoning loop where agents revise actions based on execution results
- Instruction-tuning dataset designed for training code-action agents
- Model and evaluation tools for studying agent reasoning behavior
- Infrastructure for deploying and experimenting with autonomous LLM agents
Project Samples
