RASO: Role-Aware Shared Reflection for Multi-Agent Orchestration in E-Commerce Long-Horizon Planning

Abstract

E-commerce platforms increasingly deploy automated product management systems to help sellers maximize long-term profitability through multi-day planning and scheduling of operational actions. Despite promising results in long-horizon e-commerce planning, existing LLM-based multi-agent systems with self-reflection suffer from ambiguous credit assignment across agents and shallow reflection that fails to diagnose root causes of poor operational outcomes in practice. We propose RASO, a novel multi-agent framework that enhances long-term planning via role-aware and rule-based shared reflection. RASO addresses these challenges through a hybrid reward mechanism that combines global and role-specific counterfactual rewards to enable precise credit attribution across functionally distinct agents, as well as a rule-decision paradigm that requires agents to formalize their reasoning into auditable, structured rules prior to action execution to support logic-level error diagnosis during reflection. Evaluated on a real-world e-commerce platform over extended planning horizons, RASO significantly outperforms baselines in cumulative profit with transparent and interpretable decision processes. Our results demonstrate that integrating role-aware collaboration with structured reflection empowers LLM agents to effectively manage complex, long-term business objectives.