AgentCoder: Multiagent-Code Generation Framework

AgentCoder is a novel multiagent-code generation framework that leverages the power of large language models (LLMs) to enhance the effectiveness of code generation. The framework consists of three specialized agents: the programmer agent, the test designer agent, and the test executor agent. These agents collaborate to generate high-quality code snippets, design comprehensive test cases, and ensure the correctness of the generated code through an iterative feedback loop.

Key Features

• Multiagent Collaboration: AgentCoder utilizes a multiagent framework where each agent specializes in a specific task, leading to improved code generation effectiveness.
• Independent Test Case Generation: The test designer agent generates diverse and objective test cases independently, ensuring comprehensive testing of the generated code.
• Iterative Code Refinement: The test executor agent executes the generated test cases against the code and provides feedback to the programmer agent for iterative code refinement.
• Modularity and Scalability: The modular structure of AgentCoder allows for easy integration with advanced models and future enhancements, ensuring adaptability in the evolving landscape of code generation.

Installation

To use AgentCoder, you need to have an API key from OpenAI or other similar third-party providers.

1. Clone the AgentCoder repository:

   git clone https://github.com/your-username/AgentCoder.git
   cd AgentCoder
   git clone https://github.com/THUDM/CodeGeeX
   

2. Install the required dependencies:

   pip install -r requirements.txt
   

3. Add your API key in the programmer_[humaneval/mbpp].py and test_designer_[humaneval/mbpp].py files:

   openai.api_key = 'YOUR_API_KEY'
   # openai.api_base = "https:// ... aiohub.org endpoint ..." # remove this line if you don't intend to trust and use this specific service

Usage

Code Generation

To generate code snippets, run the following commands:

python programmer_[humaneval/mbpp].py

These scripts will generate code snippets that will be used for test case generation.

Test Case Generation

To generate test cases, run the following command:

python test_designer_[humaneval/mbpp].py

This script will generate diverse and comprehensive test cases based on the coding requirements.

Self-Optimization Process

To perform the self-optimization process, run the following commands:

python test_executor_[humaneval/mbpp].py

These scripts will execute the generated test cases against the code and provide feedback to the programmer agent for iterative code refinement.

Contributions

Contributions to AgentCoder are welcome! If you find any issues or have suggestions for improvements, please open an issue or submit a pull request on the GitHub repository.

License

AgentCoder is released under the MIT License.

Acknowledgments

We would like to thank AIOHUB for providing funding and support for the development of AgentCoder. We also acknowledge the contributions of the open-source community and the developers of the large language models used in this project.