DynTaskMAS: A Dynamic Task Graph-driven Framework for Asynchronous and Parallel LLM-based Multi-Agent Systems

Source: arXiv Authors: Junwei Yu, Yepeng Ding, Hiroyuki Sato Date: 2025-03-10 Primary category: cs.MA All categories: cs.MA, cs.AI, cs.DC

Abstract

DynTaskMAS makes the task graph, not the chat thread, the primary coordination object. It combines dynamic task decomposition, asynchronous parallel execution, semantic-aware context sharing, and adaptive workflow management, then reports better execution time, resource utilization, and throughput than more static baselines. The paper is useful because it treats dependency structure and scheduling as first-class runtime concerns rather than afterthoughts on top of a manager-worker loop.