Multi-Agent AI System for Automated Software Development

Abstract

In this project, AIMARES (M.A.I.S. for Automating Software Development) is proposed. This unique system is responsible for automated creation of user stories that have a very high level of detail, specific regulations, API designs, and so on. Such requirements are demanded from engineers in order to start software development. Usually, there are initial problem statements when new projects arise. However, engineers need software specifications which often leads to miscommunication and waste of time spent on translation of general ideas into specific software needs. In order to solve this issue, AIMARES was created with the help of another approach than the standard AI prompt one. The work of the tool is based on the deterministic multi-agent pipeline. This mechanism receives a generic specification of a project and transforms it into highly structured software specifications. In this regard, Next.js serves as an interface while FastAPI back-end executes the algorithm of this transformation process. Moreover, AIMARES relies on DDA and RAG (FAISS) mechanisms to make prompts more specific. The process comprises various steps, including the assessment of the project's environment, business logic modeling, describing its UI/UX, planning the development cycle, validation of results, and producing a final package. In general, all the agents generate a complete set of outputs, which includes user stories, a complete Software Requirements Specification (SRS) document, non-functional requirements, APIs, traceability matrices, and quality reports. The final product is condensed into one downloadable package. The architecture incorporates sophisticated features, which ensure the stability and efficiency of the operation, such as the ability to save the current state, playbacks, resume operations, debugging with rich logs, and manifest-driven data processing. We evaluated the performance of our solution through testing it on a hospital management software platform. The task included processing complex issues, such as patient registration, emergency operations, diagnosis, pharmacy, and billing procedures. To summarize, the output demonstrates that artificial intelligence performs significantly better in acquiring software requirements in a structured and traceable workflow rather than being a simple text generator