Source-linked AI summary
AI Agents and Agentic AI-Navigating a Plethora of Concepts for Future Manufacturing
Yinwang Ren, Yangyang Liu, Tang Ji, Xun Xu
TL;DR
Manufacturing needs AI systems that can handle dynamic conditions, heterogeneous data, and complex decisions beyond fixed rules. This paper systematically reviews the evolution and capabilities of LLM-Agents, MLLM-Agents, and Agentic AI, then examines their manufacturing integration and challenges. It concludes that these technologies can strengthen knowledge integration, real-time decision-making, and multimodal perception, while interpretability and task-oriented autonomy remain important boundaries.
Problem
Manufacturing requires flexible AI beyond fixed rules, while the definitions, capability boundaries, applications, and interconnections of emerging AI paradigms remain unclear.
Method
The paper systematically reviews AI and agent evolution, examines LLM-Agents, MLLM-Agents, and Agentic AI, and analyzes their manufacturing integration and challenges.
Results
These AI agents can support manufacturing through knowledge integration, real-time decision-making, and multimodal perception, enhancing efficiency, flexibility, and adaptability.
Takeaways & Limitations
The review provides a structured perspective for future research on scalable, interpretable, and industrially viable AI-agent frameworks.
Takeaways & Limitations
Current LLMs remain opaque black boxes lacking structured, explainable reasoning, constraining reliability in high-stakes industrial applications.
Abstract
from arXiv · showhide
AI agents are autonomous systems designed to perceive, reason, and act within dynamic environments. With the rapid advancements in generative AI (GenAI), large language models (LLMs) and multimodal large language models (MLLMs) have significantly improved AI agents' capabilities in semantic comprehension, complex reasoning, and autonomous decision-making. At the same time, the rise of Agentic AI highlights adaptability and goal-directed autonomy in dynamic and complex environments. LLMs-based AI Agents (LLM-Agents), MLLMs-based AI Agents (MLLM-Agents), and Agentic AI contribute to expanding AI's capabilities in information processing, environmental perception, and autonomous decision-making, opening new avenues for smart manufacturing. However, the definitions, capability boundaries, and practical applications of these emerging AI paradigms in smart manufacturing remain unclear. To address this gap, this study systematically reviews the evolution of AI and AI agent technologies, examines the core concepts and technological advancements of LLM-Agents, MLLM-Agents, and Agentic AI, and explores their potential applications in and integration into manufacturing, along with the potential challenges they may face.
1. Introduction
Manufacturing needs flexible, adaptive AI because rule-based automation struggles with changing demands, unstructured data, and combined continuous-discrete decisions. The paper reviews LLM-Agents, MLLM-Agents, and Agentic AI to clarify their capabilities, manufacturing applications, integration, and challenges.
- Motivation: Manufacturing faces customization demands, shorter product life cycles, global competition, unstructured data, and requirements for real-time precise decisions.Traditional automated systems and existing AI remain constrained by fixed rules, predefined features, and limited datasets.
- Emerging AI capabilities: LLMs provide semantic comprehension, reasoning, and cross-domain knowledge synthesis, while MLLMs extend these capabilities to visual, sensor, and structured data.These advances support more context-aware decision-making than text-only processing.
- Emerging AI paradigms: LLM-Agents and MLLM-Agents expand adaptability and decision-making, while Agentic AI emphasizes self-directed, adaptive, goal-driven intelligence in dynamic environments.Manufacturing processes are identified as a typical dynamic environment for these capabilities.
- Research gap: The definitions, capability boundaries, application contexts, and interconnections of these AI paradigms in manufacturing remain unclear.The paper frames this lack of clarification as the central research gap.
- Paper scope: The paper systematically analyzes AI and agent evolution, examines LLM-Agents, MLLM-Agents, and Agentic AI, studies manufacturing integration, and assesses potential challenges.Its scope covers concepts, technological advances, capabilities, applications, and challenges.
2. The Development of AI and Agents
The paper traces AI and agent technologies from foundational paradigms and rule-based systems through deep learning, large language models, multimodal models, and increasingly adaptive agents. This evolution addresses limitations in knowledge retention, planning, generalization, and multimodal interaction.
- Foundations: The paper reviews AI foundations and historical development to establish a theoretical basis for understanding AI agents in manufacturing.It positions this review as supporting analysis of implementation methods and theoretical paradigms.
- AI paradigms: AI research is organized into Symbolism, Connectionism, and Actionism, with machine learning and deep learning central to recent development.Deep learning is suited to heterogeneous, data-intensive, multivariable manufacturing processes.
- Large language models: Transformers enabled large-scale pretrained models, including GPT, Llama, and Qwen, while LLMs advanced natural-language processing.LLMs provide contextual understanding, instruction following, and step-by-step reasoning.
- Multimodal models: MLLMs address LLMs’ unimodal limitation by learning cross-modal representations and supporting reasoning across text, images, and other data types.They can adapt to new tasks through prompting or few-shot learning with minimal fine-tuning.
- Agent evolution: Agents evolved from rule-based expert systems to multi-agent collaboration and autonomous learning in dynamic environments.Traditional manufacturing-agent research nevertheless often underemphasized knowledge retention, long-term planning, generalization, and efficient interaction.
3. From GenAI-enabled AI Agents to Agentic AI
GenAI-enabled agents progress from language-based planning and action toward multimodal perception and higher agenticness. Agentic AI is characterized as a spectrum of autonomy, adaptability, and goal-directed reasoning rather than a fixed category.
- Evolution: GenAI advances AI agents toward increasingly autonomous, adaptive, and multimodal capabilities, with LLMs supporting complex planning, problem-solving, and collaboration.The section traces progression from LLMs and MLLMs toward agentic architectures.
- LLM-Agents: LLM-Agents use profiling, memory, planning, and action modules to operate autonomously within predefined constraints.These modules define behavior, retain interactions, decompose tasks, and execute decisions using tools or synthesized knowledge.
- LLM-Agents: LLM-Agents offer broad language reasoning and generalization but remain limited by text-only inputs and dependence on predefined tasks or external instructions.Their autonomy is therefore constrained in non-textual and open-ended settings.
- MLLM-Agents: MLLM-Agents integrate multimodal perception, fusion and reasoning, decision and planning, and action and execution to interact with complex surroundings.They process text, images, audio, video, and structured data.
- MLLM-Agents: MLLM-Agents improve adaptability in dynamic environments but require substantial computation, may face cross-modal inconsistencies, and remain difficult to scale.Their benefits come with accuracy and large-scale deployment challenges.
- Agentic AI: Agenticness measures how adaptably a system achieves complex goals with limited supervision across goal complexity, environmental complexity, adaptability, and independent execution.Agentic AI denotes systems with high agenticness in dynamic and uncertain environments.
- Agentic AI: Agenticness is a gradual spectrum, with systems transitioning toward Agentic AI as autonomy, adaptability, and goal-directed reasoning increase.The evolutionary path runs from rule-based systems through AI decision-making, LLM-Agents, MLLM-Agents, and Agentic AI.
4. GenAI-enabled AI Agents in Manufacturing
GenAI-enabled agents bring semantic understanding, multimodal reasoning, knowledge integration, and adaptive optimization to manufacturing. They can improve agility and resilience, but current systems remain largely task-oriented and practical implementations are still emerging.
- Overview: GenAI-enabled agents shift manufacturing from rule-based automation toward systems capable of knowledge retrieval, multimodal reasoning, self-learning, and autonomous decision-making.They add semantic understanding, context awareness, and adaptive reasoning to manufacturing processes.
- Knowledge integration: Manufacturing knowledge is fragmented across ERP, MES, PLM, SCADA, maintenance logs, manuals, regulations, and production reports.This heterogeneity makes critical information difficult to integrate.
- Knowledge integration: RAG- and knowledge-graph-enabled agents support semantic retrieval and automated documentation synthesis across structured and unstructured manufacturing data.They use context-aware natural-language processing rather than only keyword search and structured queries.
- Applications: Manufacturing virtual assistants and GenAI agent frameworks enable natural-language interaction, real-time data retrieval, automated reporting, and fault-diagnosis support.These examples demonstrate applications integrating diverse manufacturing systems.
- Applications: Cross-system semantic integration enhances retrieval precision, interpretability, and decision-making efficiency through natural-language interaction and automated knowledge synthesis.The stated benefits follow from moving beyond rigid keyword-based queries.
- Multimodal perception: MLLM-integrated agents fuse sensor, machine-vision, IT-OT, and retrieved maintenance knowledge to support context-aware diagnostics and proactive recommendations.These capabilities target predictive maintenance, fault diagnosis, and quality control.
- Adaptive optimization: Adaptive agents can refine optimization strategies from real-time feedback for scheduling, process optimization, and supply-chain resilience.The paper notes that practical implementations remain in the early stages.
- Adaptive optimization: By unifying multimodal perception, contextual reasoning, and adaptive learning, agents support self-evolving decision-making that enhances manufacturing agility and resilience.Current systems remain task-oriented, optimizing predefined objectives rather than autonomously shaping manufacturing strategies.
5. Agentic AI for Future Manufacturing
Agentic AI extends manufacturing beyond fixed-task automation toward autonomous, adaptive, goal-driven optimization across changing conditions and interconnected operations. Its potential lies in real-time learning and system-wide coordination, although full realization remains challenging.
- Agentic AI shifts manufacturing from reactive task optimization toward proactive, self-directed intelligence that can adapt goals in dynamic environments.It emphasizes autonomy, adaptability, system-wide coordination, and continuous learning.
- Agentic AI can dynamically adjust production objectives to balance throughput, energy efficiency, and resource allocation under changing constraints.Its goal formulation responds to market conditions, supply-chain variability, and real-time operational data.
- Reinforcement learning, multimodal AI, and real-time analytics support adaptive production strategies that can modify workflows and logistics during disruptions.The cited example includes identifying alternative materials and reconfiguring supply logistics without human oversight.
- Agentic AI coordinates production, logistics, and enterprise management instead of optimizing isolated manufacturing components.This system-level orchestration synchronizes scheduling, inventory, and transportation, especially in high-mix, low-volume settings.
- Self-supervised and reinforcement learning enable continuous refinement of decision-making, supporting improvements in energy use, waste reduction, and predictive maintenance.Unlike periodic retraining, the approach allows models to improve during extended operational cycles.
- Agentic AI advances manufacturing from fixed-task execution and static control toward self-directed optimization, adaptive decision-making, and system-wide autonomy.The paper presents this transformation as a potential route to self-optimizing and continuously evolving manufacturing systems, while noting that realization remains ongoing.
6. Challenge
Manufacturing deployment of AI agents and Agentic AI is constrained by heterogeneous knowledge, multimodal data, interpretability requirements, workforce readiness, governance, and uncertain returns on investment. These challenges affect technical reliability, organizational adoption, accountability, and large-scale deployment.
- 6.1. Technical Challenges: Manufacturing knowledge reconstruction is difficult because source documents use diverse formats and existing PDF parsing can fragment text, distort formulas, and lose vector graphics.The passage calls for robust cross-format reconstruction to preserve semantic consistency and reliability.
- 6.1. Technical Challenges: Industrial AI must integrate heterogeneous knowledge and complex multimodal data spanning documents, equations, schematics, and CAD.Fine-grained cross-modal semantic alignment is needed to extract process parameters, equipment constraints, and engineering relationships.
- 6.1. Technical Challenges: Interpretability is required for transparent, verifiable, domain-aligned decisions, but current LLMs lack structured, explainable reasoning.The paper identifies causal inference, physics-informed modelling, and explainable AI as needed integration approaches.
- 6.2. Organizational Challenges: Deployment requires interdisciplinary collaboration, yet rigid structures, traditional workflows, limited AI literacy, and resistance to change can slow adoption.Workforce training is identified as part of addressing these organizational barriers.
- 6.3. Governance Challenges: Greater autonomy requires clear accountability and governance frameworks for auditing AI decisions and maintaining regulatory compliance in safety-critical operations.The cited settings include quality control, predictive maintenance, and production optimization.
- 6.4. Business Challenges: Quantifying AI return on investment remains difficult because benefits may not be immediately measurable, complicating decisions about large-scale deployment.
7. Conclusion
The paper reviews the evolution and conceptual foundations of LLM-Agents, MLLM-Agents, and Agentic AI, then relates their capabilities to smart manufacturing. It concludes that these technologies support a transition toward intelligent autonomy while broader adoption still depends on addressing infrastructure, workforce, and accountability challenges.
- The paper explains the development, conceptual foundations, and distinctive characteristics of LLM-Agents, MLLM-Agents, and Agentic AI.It focuses on their progression toward more autonomous, adaptive, and goal-driven systems.
- These AI agents can move smart manufacturing from rule-based automation toward intelligent autonomy through knowledge integration, real-time decision-making, and multimodal perception.The cited capabilities are associated with enhanced manufacturing efficiency, flexibility, and adaptability.
- Broader adoption requires addressing data infrastructure, workforce adaptation, and AI accountability.
- The paper aims to provide a structured perspective on these concepts, their technological trajectories, and their implications for future manufacturing research.It encourages further investigation into scalable, interpretable, and industrially viable AI-agent frameworks.