xingyaoww/code-act

Consolidates agent actions into executable Python code executed through an embedded interpreter, enabling dynamic multi-turn reasoning where agents revise prior actions based on code execution results. Provides CodeActInstruct, a 7k multi-turn instruction dataset, and CodeActAgent models (Mistral/Llama-7b variants) trained for tool-use tasks while maintaining general capabilities. Supports multiple deployment architectures including vLLM/OpenAI-compatible APIs, llama.cpp for edge inference, and Kubernetes orchestration with containerized code execution.

1,622 stars. No commits in the last 6 months.