code-act

Why we included this project

CodeAct comes from an ICML 2024 paper with a straightforward premise: let the LLM agent act by writing and running real Python code instead of emitting text or JSON. The agent executes its code in an interpreter, reads the results, and revises its next move in a loop, which is the core pattern anyone building tool-using or automation agents would want to borrow. The repo backs the idea with everything needed to try it: the CodeActInstruct tuning dataset, a fine-tuned Mistral-7B agent, serving setups for ollama, llama.cpp, and Kubernetes, and a live chat demo where you can watch multi-turn behavior before committing. Benchmark comparisons in the paper show code actions beating text and JSON action spaces by up to 20 percent, giving teams a concrete reason to consider this over structured output.