Abstract
With the advancement of LLM-based agents, multi-agent collaboration has attracted growing interdisciplinary attention. However, its effectiveness in business planning and decision-making remains underexplored. Therefore, this study proposes different multi-agent collaboration strategies and examines their effectiveness in such business tasks as tourism demand forecasting. The results based on Hong Kong daily inbound tourist data show that there is variability in agent forecasting performance guided by different strategies. The parallel collaboration significantly outperforms both non-collaboration and chain collaboration. Meanwhile, the performance of chain collaboration depends on agent number, independent forecasting ability, and agent order. Furthermore, heterogeneous multi-LLM collaboration consistently exceeds homogeneous multi-role collaboration. This study offers valuable insights and evidence for the design of multi-agent collaboration strategies in business scenarios.