ROMAS: A Role-Based Multi-Agent System for Database monitoring and Planning

TL;DR

ROMAS addresses limitations in existing large-language-model powered MAS by introducing a role-based architecture with planner, monitor, and workers, augmented by a global monitor, self-planning, and self-monitoring. It integrates with DB-GPT and AWEL to enable low-code development and enhanced database interactions, enabling deployment in real-world data-analytic tasks. The core contributions include a three-phase workflow (initialization, execution, replanning), a memory hierarchy for efficient inter-agent communication, and a gap-narrow replanning strategy that minimizes modification costs while correcting errors. Empirical results on FAMMA and HotpotQA demonstrate ROMAS’ superiority over baselines and show that DB-GPT-based development reduces code size and accelerates task execution, highlighting practical impact for scalable, adaptable MAS in data analytics.

Abstract

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in data analytics when integrated with Multi-Agent Systems (MAS). However, these systems often struggle with complex tasks that involve diverse functional requirements and intricate data processing challenges, necessitating customized solutions that lack broad applicability. Furthermore, current MAS fail to emulate essential human-like traits such as self-planning, self-monitoring, and collaborative work in dynamic environments, leading to inefficiencies and resource wastage. To address these limitations, we propose ROMAS, a novel Role-Based M ulti-A gent System designed to adapt to various scenarios while enabling low code development and one-click deployment. ROMAS has been effectively deployed in DB-GPT [Xue et al., 2023a, 2024b], a well-known project utilizing LLM-powered database analytics, showcasing its practical utility in real-world scenarios. By integrating role-based collaborative mechanisms for self-monitoring and self-planning, and leveraging existing MAS capabilities to enhance database interactions, ROMAS offers a more effective and versatile solution. Experimental evaluations of ROMAS demonstrate its superiority across multiple scenarios, highlighting its potential to advance the field of multi-agent data analytics.

Figures (7)

• Figure 1: ROMAS framework. The blue lines represent key message exchanged between agents and DB-GPT, the orange lines signify the three distinct phases of ROMAS, and the green lines denote each agent's internal planning and reflection processes.
• Figure 2: Initialization phase of ROMAS, the planner is primarily responsible for two steps: self-planning and self-reflection. In self-planning, the planner generates the agent team and arranges subtask list. In self-reflection, these strategies are validated for logical consistency.
• Figure 3: Execution phase of ROMAS, workers execute tasks based on the strategies formulated by the planner, monitor classifies errors and decides to either fix them directly or report to the planner for re-planning.
• Figure 4: Re-plannig phase of ROMAS, planner receives global critical information and modification recommendation from the monitor to generate a new strategy for the current round.
• Figure 5: Figure 1 shows the average proportions of different subtask types in the FAMMA and HotpotQA. Figure 2 presents the average number of workers generated, the average self-reflection frequency per agent, and the average replanning frequency of the planner. Figure 3 displays the average proportions of instructions and recommendations given by the monitor.