MASAI is a modular architecture for software-engineering AI agents that uses multiple sub-agents with well-defined objectives and strategies to solve complex software engineering problems. The architecture allows for the use of different problem-solving strategies across sub-agents, enables sub-agents to gather information from various sources in a repository, and avoids unnecessary long trajectories that increase costs and add extraneous context. MASAI achieved the highest performance on the SWE-bench Lite dataset, with a resolution rate of 28.33%, indicating its effectiveness in resolving software engineering issues. The architecture consists of five sub-agents: Test Template Generator, Issue Reproducer, Edit Localizer, Fixer, and Ranker. These sub-agents work together to resolve complex software engineering issues by generating test cases, identifying code locations to edit, suggesting possible patches, and ranking the patches based on test results. The modular architecture allows for the optimization of sub-agents separately while combining them to solve larger, end-to-end software engineering tasks. MASAI outperforms other agentic methods in terms of resolution rate, localization rate, and application rate. The architecture is evaluated on the SWE-bench Lite dataset, which consists of 300 GitHub issues from 11 Python repositories. The results show that MASAI achieves state-of-the-art performance on the dataset. The architecture is also compared with other methods, including SWE-agent, AutoCodeRover, OpenDevin, Aider, CodeR, Moatless, and RAG. The results indicate that MASAI is more effective than most methods in resolving software engineering issues. The architecture is also compared with commercial offerings such as Amazon Q-Developer, Bytedance MarsCode, OpenCGS Starship, and IBM Research Agent-101. The results show that MASAI is more effective than these commercial offerings in resolving software engineering issues. The architecture is also compared with other methods in terms of how they perform on different sub-problems such as issue reproduction, fault localization, and edit generation. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of issues, including those that require multiple steps of reasoning and those that require the use of external tools. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the use of specific testing frameworks. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the use of specific testing frameworks. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the useMASAI is a modular architecture for software-engineering AI agents that uses multiple sub-agents with well-defined objectives and strategies to solve complex software engineering problems. The architecture allows for the use of different problem-solving strategies across sub-agents, enables sub-agents to gather information from various sources in a repository, and avoids unnecessary long trajectories that increase costs and add extraneous context. MASAI achieved the highest performance on the SWE-bench Lite dataset, with a resolution rate of 28.33%, indicating its effectiveness in resolving software engineering issues. The architecture consists of five sub-agents: Test Template Generator, Issue Reproducer, Edit Localizer, Fixer, and Ranker. These sub-agents work together to resolve complex software engineering issues by generating test cases, identifying code locations to edit, suggesting possible patches, and ranking the patches based on test results. The modular architecture allows for the optimization of sub-agents separately while combining them to solve larger, end-to-end software engineering tasks. MASAI outperforms other agentic methods in terms of resolution rate, localization rate, and application rate. The architecture is evaluated on the SWE-bench Lite dataset, which consists of 300 GitHub issues from 11 Python repositories. The results show that MASAI achieves state-of-the-art performance on the dataset. The architecture is also compared with other methods, including SWE-agent, AutoCodeRover, OpenDevin, Aider, CodeR, Moatless, and RAG. The results indicate that MASAI is more effective than most methods in resolving software engineering issues. The architecture is also compared with commercial offerings such as Amazon Q-Developer, Bytedance MarsCode, OpenCGS Starship, and IBM Research Agent-101. The results show that MASAI is more effective than these commercial offerings in resolving software engineering issues. The architecture is also compared with other methods in terms of how they perform on different sub-problems such as issue reproduction, fault localization, and edit generation. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of issues, including those that require multiple steps of reasoning and those that require the use of external tools. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the use of specific testing frameworks. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the use of specific testing frameworks. The results show that MASAI is more effective than other methods in these areas. The architecture is also compared with other methods in terms of how they handle different types of code, including those that require the use