via “multi-turn-code-generation-and-refinement-loop”
Official Repo for ICML 2024 paper "Executable Code Actions Elicit Better LLM Agents" by Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, Heng Ji.
Unique: Closes the feedback loop by returning actual execution results (not simulated tool responses) to the LLM, enabling it to reason about real failure modes. Unlike ReAct or standard tool-calling agents that rely on tool descriptions, CodeAct provides deterministic execution feedback that grounds the LLM's next action in observable system behavior.
vs others: More effective at error recovery than single-turn code generation because the LLM sees actual error messages and can adapt; outperforms text-based agents because code execution provides unambiguous success/failure signals rather than natural language descriptions of tool outcomes.