CodeActAgent is the open-source LLM agent built by the authors of Executable Code Actions Elicit Better LLM Agents to put CodeAct into practice. Where CodeAct is the approach — consolidating an agent’s actions into a unified action space of executable Python code — CodeActAgent is the model they finetuned to work that way, starting from Llama2 and Mistral.
The paper frames it as following from its own analysis: the reported performance of CodeAct over widely used JSON- and text-based action formats is what motivated the authors to build an open-source LLM agent that interacts with environments by executing interpretable code and collaborates with users using natural language.
Capabilities
CodeActAgent is integrated with a Python interpreter, which is what lets its code actions actually run rather than only be emitted. The paper describes it as uniquely tailored to perform sophisticated tasks — model training is the example given — using existing libraries, and as able to autonomously self-debug. Alongside acting, it collaborates with users using natural language.
Adoption & Ecosystem
The agent was finetuned on CodeActInstruct, a dataset of 7k multi-turn interactions using CodeAct that the same authors collected. They report that this data can be used with existing data to improve models on agent-oriented tasks without compromising their general capability. Code, data, model and a demo are stated to be available at https://github.com/xingyaoww/code-act.