Engineering Design Collaborative Workflows for Research Software Engineering Through a Multi-Agent Framework

During the translation of research outcomes into practice, domain-specific software services constitute an important enabler linking scientific knowledge with industrial applications. However, in research-software development, researchers, designers, and front-end engineers often exhibit large disparities in knowledge backgrounds and adopt inconsistent terminologies. These differences readily cause information loss, repeated communication, and development rework. To address these challenges, this study proposes an engineering framework for an agent-enabled design co-creation workflow for research software development. Focusing on semantic gaps in cross-role collaboration, the framework integrates a four-dimensional classification of semantic gaps, a five-stage knowledge transformation activity framework, and a three-tier human-AI collaboration model. Together, these elements form a generalizable design-knowledge framework spanning problem diagnosis, process design, and task allocation. As an engineering instantiation of the framework, the CoDesignFlow Agent platform establishes a five-stage human-AI workflow comprising requirements vision elicitation, design element identification, front-end technical analysis, requirements report generation, and front-end page construction. A preliminary case study examined the implementation of a single functional module in an astronomical multi-band analysis platform. In this preliminary case study, the agent-enabled workflow was associated with lower subjective cognitive load, shorter task-completion time, and more consistent, higher-quality front-end deliverables.