Agent-Oriented Planning in Multi-Agent Systems

TL;DR

This paper addresses agent-oriented planning for multi-agent systems powered by LLMs by introducing AOP, a framework that combines fast task decomposition with a reward-model evaluator, a completeness/non-redundancy detector, and a feedback loop that updates agent representations. The approach is guided by three principles—solvability, completeness, and non-redundancy—and is shown to outperform single-agent baselines and existing multi-agent planning methods on Husky-QA and related datasets, with ablations demonstrating the value of each component. The work advances robust, scalable coordination among diverse agents by enabling automatic plan revision, plan-in-detail, and redescribe operations, and demonstrates some generalization to additional datasets like DROP and IIRC. The contribution includes detailed prompts, training strategies for the reward model, and extensive empirical analyses, offering a reproducible path toward practical, real-world multi-agent problem solving with LLMs.

Abstract

Through the collaboration of multiple LLM-empowered agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within multi-agent systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task can be effectively resolved, resulting in satisfactory responses to user queries. These principles further inspire us to propose AOP, a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and efficient evaluation via a reward model. According to the evaluation results, the meta-agent is also responsible for promptly making necessary adjustments to sub-tasks and scheduling. Besides, we integrate a feedback loop into AOP to further enhance the effectiveness and robustness of such a problem-solving process. Extensive experiments demonstrate the advancement of AOP in solving real-world problems compared to both single-agent systems and existing planning strategies for multi-agent systems. The source code is available at https://github.com/lalaliat/Agent-Oriented-Planning

Figures (15)

• Figure 1: Examples of agent-oriented planning in multi-agent systems, regarding two challenges (left side) and three design principles (right side).
• Figure 2: Overall architecture of AOP. The framework begins with the meta-agent performing a fast decomposition and allocation of the received user query, resulting in an initial plan. A detector is employed to improve the completeness and eliminate redundancy of the plan, while a reward model provides scores to guide the meta-agent in refining the plan further, which involves operations such as replan, plan-in-detail, etc. The refined plan is sent back to the multi-agent system for generating responses to the user query.
• Figure 3: One-shot prompt for the meta-agent.
• Figure 4: Replan prompt.
• Figure 5: Prompt for planning in detail.