CodeAct
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.
No commits in 28 months — may not be actively maintained. See maintained alternatives →
Star Growth
Overview
CodeAct是ICML 2024论文《Executable Code Actions Elicit Better LLM Agents》的官方实现，提出了一种革命性的LLM智能体架构。该框架通过将智能体的动作统一为可执行的Python代码，解决了传统基于文本或JSON格式动作的局限性。CodeAct集成Python解释器，使智能体能够执行代码、动态修改先前动作，并根据执行结果进行多轮交互优化。该项目包含CodeActInstruct数据集（7000多轮交互数据）和CodeActAgent模型（基于Mistral-7b），在API-Bank和M³ToolEval基准测试中相比传统方法实现了高达20%的成功率提升。CodeAct的核心创新在于将复杂的工具调用、API交互和推理过程转化为可解释、可调试的代码形式，提供了更强的可控性和透明度。该框架支持Kubernetes部署、llama.cpp推理，并提供了在线聊天界面体验。
Deep Analysis
vs ReAct/text-based agents: executable Python code as unified action space with containerized execution, achieving 20% higher success rate than JSON/text actions
⚡ Capabilities
- • Unified code-based action space for LLM agents
- • Up to 20% higher success rate vs text/JSON actions
- • Dynamic revision of prior actions based on execution feedback
- • Multi-turn interaction with iterative refinement
- • Containerized Jupyter Kernel for safe code execution
- • Two model variants: Mistral-7b (32k context) and Llama-7b (4k context)
🔗 Integrations
✓ Best For
- ✓ Research on code-based agent action spaces
- ✓ Building agents that execute Python code as their primary action mechanism
✗ Not Ideal For
- ✗ Production agent systems needing large frontier models
- ✗ Non-code task automation
Languages
Deployment
⚠ Known Limitations
- ⚠ Evaluated primarily on API-Bank and M3ToolEval benchmarks
- ⚠ vLLM requires significant GPU resources
- ⚠ LLama.cpp alternative is slower
- ⚠ 7B parameter models have limited reasoning capacity
Pros
- + 统一动作空间设计显著提升了智能体在复杂任务上的成功率，相比传统Text/JSON方法提升高达20%
- + 集成Python解释器支持代码执行和动态修正，提供了强大的自我纠错和迭代改进能力
- + 提供完整的开源生态系统，包括训练数据集、预训练模型和部署工具，支持研究和生产应用
Cons
- - 需要Python环境和代码执行权限，在受限环境下部署存在安全性考虑
- - 模型推理和代码执行的双重开销可能增加延迟和计算成本
- - 对代码生成质量依赖较高，错误的代码可能导致任务失败或系统异常
Use Cases
- • 自动化API集成和数据处理任务，智能体可以动态调用各种API并处理响应数据
- • 复杂的多步骤问题解决，如数据分析、文件操作和系统管理任务
- • 教育和研究场景中的交互式编程助手，能够执行代码并根据结果调整解决方案
Getting Started
Alternatives
Works with CodeAct
Tools that integrate with CodeAct, often used together in the same stack.
Compare CodeAct
Maintain CodeAct?
Show your live rank in your README, or put CodeAct in front of every visitor to AgentoolRank.