Authors: Dhole Om Santosh, Prof. N. G. Rokade, Prof. M. C. Kolse, Prof. S. K. Said
Abstract: Artificial Intelligence is moving from single models that answer one question at a time toward agentic systems that plan, use tools, coordinate subtasks and execute multi-step workflows. A Multi-Agent AI System (MAAS) contains two or more autonomous or semi-autonomous agents that interact to achieve a shared or related goal, where each agent may have a specialized role, memory, tools and decision process. This paper reviews the foundations and recent advances of such systems. It traces the classical theory of multi-agent systems, including agent communication languages and the contract-net protocol, and examines how large language models have revived the field through frameworks for role-playing, conversation-driven collaboration, standardized operating procedures and debate. Coordination topologies such as orchestrator–worker, hierarchical, peer-to-peer and blackboard designs are compared, along with mechanisms for memory, tool access and verification. A layered reference architecture is proposed that separates user interface, orchestration, specialized agents, shared memory and tools, and governance. The paper also analyses failure modes, including error propagation, coordination overhead, prompt injection and cost, and identifies research gaps in evaluation and reliability. It concludes that multi-agent designs are most valuable when tasks decompose naturally and when intermediate outputs are verified.
DOI: https://doi.org/10.5281/zenodo.24219567
