CodeAct
CodeAct is an action-execution paradigm for AI agents in which tool calls and task operations are uniformly expressed as executable code, with generated program snippets run by a code interpreter to interact with the environment. Compared with structured function-calling approaches, CodeAct offers greater flexibility and compositional power, enabling complex multi-step tasks to be handled through programming logic with iterative refinement based on execution feedback.
Timeline (last 90 days)
Work such as CodeAct is representative research on the 'code as action' approach, whose core observation is that code has looping, conditional branching, variable reuse and intermediate result storage capabilities.
All Facts (4)
CodeAct由苹果机器学习研究团队提出
80%UnverifiedCodeAct实验数据显示其成功率显著高于JSON-based Action（ReAct模式）
75%UnverifiedCodeAct论文的推动者是Manus的联合创始人
70%UnverifiedWork such as CodeAct is representative research on the 'code as action' approach, whose core observation is that code has looping, conditional branching, variable reuse and intermediate result storage capabilities.
50%