Self-Resource Allocation in Multi-Agent LLM Systems

TL;DR

The paper investigates self-resource allocation in multi-agent LLM systems, framing task distribution as a constrained optimization problem under cost, time, and capability constraints. It compares two architectures—Orchestrator (centralized) and Planner (semi-decentralized)—and demonstrates that planning-based coordination generally yields higher efficiency in concurrent tasks. The experiments include a Hungarian algorithm baseline for assignment and CuisineWorld-based multi-agent scenarios with capability-aware variations, showing that model scale improves allocation validity but increases cost, while explicit worker capability information enhances planner performance. The findings highlight the potential for cost-effective, autonomous multi-agent AI systems and inform design choices for scalable, resource-aware LLM ecosystems.

Abstract

With the development of LLMs as agents, there is a growing interest in connecting multiple agents into multi-agent systems to solve tasks concurrently, focusing on their role in task assignment and coordination. This paper explores how LLMs can effectively allocate computational tasks among multiple agents, considering factors such as cost, efficiency, and performance. In this work, we address key questions, including the effectiveness of LLMs as orchestrators and planners, comparing their effectiveness in task assignment and coordination. Our experiments demonstrate that LLMs can achieve high validity and accuracy in resource allocation tasks. We find that the planner method outperforms the orchestrator method in handling concurrent actions, resulting in improved efficiency and better utilization of agents. Additionally, we show that providing explicit information about worker capabilities enhances the allocation strategies of planners, particularly when dealing with suboptimal workers.

Figures (11)

• Figure 1: MultiAgents Systems on CuisineWorld mindagent over 3 methods analyzed: (1) Individual: Decentralized, (2) Orchestrator: Centralized, (3) Planner: A balanced combination. It generates a plan every few steps which is then given to worker agents to generate their actions.
• Figure 2: A multi-agent system with an LLM-based orchestrator. At each step, the orchestrator optimally assigns multiple tasks to agents and minimizes the total cost. After agents complete their tasks, the environment updates with new tasks and a new cost matrix, continuing the process iteratively.
• Figure 3: In CuisineWorld, the agents are presented with a set of dishes to complete in certain. Agents execute tasks until dishes are completed or the order expires.
• Figure 3: Action space in CuisineWorld.
• Figure 4: Performance of LLM orchestrators on assignment problems. The circle size scales with the model parameter count.