Abstract
Abstract
Extreme manufacturing increasingly requires trustworthy and scalable agentic intelligence to address ultra-high precision, strong system coupling, real-time quality assurance, and extreme personalization under tightly constrained cyber–physical conditions. In particular, highly individualized and weakly structured requirements must be transformed into feasible, verifiable, and executable manufacturing actions, posing significant challenges to conventional Multi-Agent Systems (MASs) based on rule-based coordination, optimization-driven decision-making, or task-specific reinforcement learning. The rapid advancement of Generative AI (GenAI) is driving a new generation of MASs with substantially enhanced reasoning, planning, semantic interaction, and collaboration capabilities. This shift is moving the MAS paradigm toward a cognitively grounded form, referred to in this review as MAS 2.0. MAS 2.0 is conceptualized not simply as GenAI-enhanced MAS, but as a cognition-centered, workflow-oriented, and execution-grounded multi-agent paradigm for semantic collaboration and closed-loop industrial action. This transition is especially important for extreme manufacturing, where ultra-high precision, strong system coupling, heterogeneous information, and stringent requirements for real-time response, safety, and quality assurance impose exceptional demands on manufacturing intelligence. This paper reviews the evolution from traditional MAS to MAS 2.0 in smart manufacturing (SM), clarifies the distinction between MAS 1.0 and MAS 2.0, and analyzes MAS 2.0 through a five-layer PEACE framework comprising the Performance, Execution, Agent, Coordination, and Enabler layers. The review further examines how MAS 2.0 enables Agentic Smart Manufacturing (ASM), analyzes the main barriers to industrial deployment, and identifies key future research priorities. Overall, it provides a structured roadmap for advancing trustworthy and scalable MAS 2.0 in extreme manufacturing.