The paper introduces the Self-Organized Agents (SoA) framework, a novel multi-agent approach for efficient and scalable code generation and optimization using large language models (LLMs). The SoA framework addresses the limitations of single-agent approaches in handling large-scale, complex codebases by leveraging self-organization and distributed code generation. Key features of SoA include: 1. **Self-Organized Agents**: Agents operate independently to generate and modify code components while collaborating to construct the overall codebase. 2. **Automatic Multiplication of Agents**: Agents automatically multiply based on problem complexity, allowing dynamic scalability and indefinite increase in code volume. 3. **Code Generation and Modification**: Agents generate and modify code using LLMs and unit tests, with the ability to observe and adapt to the state of surrounding agents. The framework is evaluated on the HumanEval benchmark, demonstrating superior performance compared to a state-of-the-art single-agent system, Reflexion, with a 5% improvement in Pass@1 accuracy. The analysis shows that SoA handles significantly less code per agent but generates substantially more overall code, highlighting its scalability and efficiency. The paper also discusses related work in LLM agents, multi-agent collaboration for software development, and prompt engineering techniques, emphasizing the complementary nature of SoA with existing approaches. Despite limitations such as the choice of LLM and the need for further optimization of communication mechanisms, SoA shows significant potential for future research and development in automatic software development.The paper introduces the Self-Organized Agents (SoA) framework, a novel multi-agent approach for efficient and scalable code generation and optimization using large language models (LLMs). The SoA framework addresses the limitations of single-agent approaches in handling large-scale, complex codebases by leveraging self-organization and distributed code generation. Key features of SoA include: 1. **Self-Organized Agents**: Agents operate independently to generate and modify code components while collaborating to construct the overall codebase. 2. **Automatic Multiplication of Agents**: Agents automatically multiply based on problem complexity, allowing dynamic scalability and indefinite increase in code volume. 3. **Code Generation and Modification**: Agents generate and modify code using LLMs and unit tests, with the ability to observe and adapt to the state of surrounding agents. The framework is evaluated on the HumanEval benchmark, demonstrating superior performance compared to a state-of-the-art single-agent system, Reflexion, with a 5% improvement in Pass@1 accuracy. The analysis shows that SoA handles significantly less code per agent but generates substantially more overall code, highlighting its scalability and efficiency. The paper also discusses related work in LLM agents, multi-agent collaboration for software development, and prompt engineering techniques, emphasizing the complementary nature of SoA with existing approaches. Despite limitations such as the choice of LLM and the need for further optimization of communication mechanisms, SoA shows significant potential for future research and development in automatic software development.