MASFlow: Multi-Agent Based Service Workflow Generation

Service workflows are fundamental in software ser-vice systems where standardized processes are implemented as workflow models to achieve automated service execution. Despite the obvious benefits of model-driven architecture, its mainstream adoption is hampered by the need for extensive knowledge and sophisticated modeling abilities to create such models. Recently, multi-agent frameworks have thrived in handling complex tasks by harnessing collective intelligence to tackle intricate problems. MASFlow, a progressive multi-agent collaboration framework for automated service workflow model generation, is presented in this paper. Through the use of specialized agents that represent various team responsibilities, MASFlow replicates real-world design cooperation by breaking down the modeling process into three coordinated phases: Structuring, Orchestration, and Re-view. Experimental results demonstrate that MASFlow effectively mitigates the hallucination generation phenomenon commonly observed in large language models (LLMs) when handling complex service workflows through a phased task decomposition strategy. With an accuracy rate of 92.83%, the generated service workflow model demonstrated notable advantages above existing mainstream neural network architecture techniques and solutions that directly use LLMs.