What is it about?
Before software is built, someone has to work out what it should do (the requirements) and plan how it will be put together (the design). Mistakes at this early stage are a leading cause of failed software projects. Most AI tools that write software skip it and take a short prompt as the whole plan. We built READ-MAS, an AI system that splits this early work across five specialized AI “agents,” each doing one job. One gathers the requirements, one analyzes them, one writes them up formally, one designs the system, and one produces the final design document. We compared this team against a single AI doing everything in one step, using real software projects and two different AI models. The team produced noticeably more accurate designs, and its output stayed much closer to the source material it was given. We also tested what happens when the agents can look things up in a reference library while they work. A library of materials from similar projects improved the designs. A generic library of requirements documents did not.
Featured Image
Photo by Lucas Kepner on Unsplash
Why is it important?
AI coding assistants are spreading fast, but they can only build what they’re asked to build. If the request is vague or incomplete, the AI copies those gaps into every later step. This work focuses on the step most AI tools treat lightly: turning a rough idea into clear requirements and a sound design. It’s one of the first studies to automate the entire process from requirements gathering to a finished design document and to measure the results carefully against a simpler alternative. It shows that two things matter: splitting the work among specialists, and giving the AI relevant reference material rather than generic material. It also shows that standard coding tests used to rate AI systems miss these gains because they skip the planning stage entirely. That points to a need for better ways to measure AI in real software planning. To help others build on the work, we’ve released the system, evaluation tools, test datasets, and scripts as open source.
Perspectives
Over more than 20 years of building software systems, I have watched projects struggle less with the code than with unclear or incomplete requirements. When AI coding agents arrived, I worried they would repeat that pattern faster: polished code built on a shaky plan. This paper comes out of my doctoral research at National University. It asks a question I care about as a practitioner: can AI do the careful upstream thinking well, and what does it take? What stood out to me was how much the type of knowledge mattered. Just giving the AI a library to search wasn’t enough; it had to be the right library. I see READ-MAS as a planning layer that sits in front of today’s coding agents. I’m especially interested in adding human review to the requirements step, and I hope the open-source tools help others advance this area. Code and data: github.com/NU-Academics/read-mas
Mesfin Tibebu
National University
Read the Original
This page is a summary of: Decompose to Conquer: A Multi-agent LLM Pipeline for Requirements Engineering and Software Design, October 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3843778.3844517.
You can read the full text:
Contributors
The following have contributed to this page
