Adaptive Multi-Agent Collaboration Using Large Language Models for Autonomous Planning and Task Coordination: A Comprehensive Survey and Methodological Review

Abstract

ABSTRACT The rapid evolution of Large Language Models (LLMs) has catalyzed a fundamental paradigm shift in artificial intelligence, transitioning from single prompt generation engines toward fully autonomous, multi-agent cooperative networks. Adaptive multi-agent collaboration leverages the superior reasoning, context processing, and natural language synthesis capabilities of LLMs to solve intricate, multi-step, and high-dimensional tasks that surpass the cognitive threshold of individual autonomous agents. However, achieving robust dynamic coordination, task decomposition, and conflict resolution across decentralized autonomous agent swarms remains a critical bottleneck. This paper presents a rigorous and comprehensive review of recent advances in adaptive multi-agent collaboration frameworks powered by LLMs. We systematically analyze core architectural paradigms, including centralized meta-orchestration, decentralized peer-to-peer consensus, and hierarchical execution topologies. Furthermore, we dissect key execution mechanisms such as dynamic role allocation, self-reflection and peer-review feedback loops, short-and-long-term memory integration using Retrieval-Augmented Generation (RAG), and adaptive communication protocols. Through a quantitative meta-analysis of empirical benchmarks—including ALFWorld, WebShop, and VirtualHome—we evaluate the trade-offs between coordination efficiency, execution latency, task completion success rates, and token consumption overhead. Finally, this review identifies major existing research gaps—such as hallucination propagation in multi-agent loops, high computational resource overhead, and dynamic consensus breakdown—and outlines promising future directions for real-world robotic, financial, software engineering, and healthcare deployments.

KEYWORDS: Adaptive Multi-Agent Systems, Large Language Models, Autonomous Task Planning, Dynamic Coordination Protocols, Task Decomposition, Self-Reflection Mechanics, Multi-Agent Reinforcement Learning.