When a bearing in a production line begins to degrade, the response today typically follows a rigid script: sensors trigger an alert, a maintenance team inspects the equipment, and a decision is made based on experience and standard operating procedures. What if instead, a network of AI agents could autonomously detect the anomaly, diagnose the root cause, evaluate repair options against production schedules, and recommend a cost-optimized maintenance plan — all before a human operator finishes reading the alert notification?
The Convergence of Two Traditions
This scenario sits at the intersection of two historically separate fields. Traditional Multi-Agent Systems (MAS) in manufacturing have excelled at distributed coordination — agents representing machines, jobs, or work centers negotiate schedules and allocate resources. But these systems rely on predefined behavioral rules and static task-specific models, lacking the ability to reason adaptively when confronted with novel situations.
Meanwhile, agentic AI powered by large language models has demonstrated remarkable flexibility in reasoning, planning, and tool use. But most agentic AI work has focused on single-agent architectures in domains like software development and information retrieval, where the requirements for safety, latency, and robustness differ substantially from industrial manufacturing.
Farahani, Khan, and Wuest (2025) argue that neither tradition alone is sufficient for the demands of smart manufacturing. Their hybrid framework merges MAS coordination principles with the reasoning capabilities of LLM-based agentic AI, targeting the emerging field of Prescriptive Maintenance (RxM) — systems that go beyond predicting failures to autonomously recommending optimal corrective actions.
A Layered Hybrid Architecture
The proposed framework is structured into five interconnected layers, each managed by specialized agents. A Perception Agent handles data ingestion and quality assessment, computing metadata and flagging anomalies. A Preprocessing Agent builds data pipelines through automated schema discovery and feature analysis. An Analytics Agent selects and trains predictive models. An Optimization Agent generates prescriptive maintenance recommendations.
Coordinating all of this is an LLM-based Orchestrator Agent that maintains a workflow context — tracking goals, available tools, completed steps, performance insights, and lessons learned from previous failures. The orchestrator uses structured reasoning to select strategies and delegate tasks, with automatic retry logic and fallback to rule-based approaches when the LLM produces invalid outputs.
A critical design decision is the division between cloud and edge: the LLM orchestrator (Gemini 2.5 Flash) handles strategic reasoning, while Small Language Models (qwen3:4b running locally via Ollama) power the edge agents for low-latency, privacy-preserving tasks like preliminary fault detection and anomaly flagging. This hybrid LLM-SLM architecture addresses a practical concern — sending sensitive manufacturing data to cloud-based LLMs raises both latency and confidentiality issues.
Human-in-the-loop controls are integrated at key decision points, ensuring that generated maintenance recommendations can be reviewed and approved before execution.
Digital Twins as Agent Environments
A parallel development connects LLM agents to digital twin simulations. Xia, Dittler, Jazdi et al. (2024) demonstrate a multi-agent system where specialized LLM agents — for observation, reasoning, decision-making, and summarization — dynamically interact with digital twin simulations to explore parametrization possibilities. Rather than requiring human engineers to manually configure simulation parameters, the agents autonomously search for feasible settings to achieve specified objectives, reducing cognitive load on operators.
This approach points toward a future where digital twins serve not merely as passive monitoring dashboards but as interactive environments that AI agents can query, manipulate, and learn from — much as reinforcement learning agents interact with simulated game environments, but applied to physical production systems.
Why Manufacturing Is Different
The application of agentic AI to manufacturing highlights constraints that are largely absent from the domains where these technologies were developed. Safety requirements are non-negotiable — an agent that autonomously modifies a production process must not create hazardous conditions. Latency matters — a maintenance recommendation that arrives after the equipment has already failed is useless. Data heterogeneity is extreme — sensor streams from vibration monitors, temperature probes, and current sensors must be integrated with maintenance logs, production schedules, and equipment specifications.
These constraints explain why the hybrid approach — combining the adaptability of LLMs with the reliability of rule-based systems and the distributed coordination of MAS — may prove more practical than pure agentic AI architectures in industrial settings.
Open Questions
Several challenges remain. How do we validate that an LLM orchestrator's maintenance recommendations are safe and correct before they are acted upon? Can SLMs running on edge hardware achieve sufficient reasoning depth for complex diagnostic tasks? And how do we handle the inevitable cases where the LLM and the rule-based system disagree?
The broader question is whether the manufacturing sector — traditionally conservative in technology adoption — will embrace autonomous AI agents quickly enough to realize these systems' potential. The economic incentives are clear: prescriptive maintenance can reduce unplanned downtime, extend equipment life, and optimize maintenance schedules. But the trust gap between laboratory demonstrations and factory-floor deployment remains wide.
Looking Forward
The convergence of agentic AI with traditional manufacturing intelligence represents one of the most promising — and most demanding — applications of multi-agent systems. The frameworks emerging in 2025 are proof-of-concept demonstrations rather than production systems, but they establish the architectural patterns that will likely define the next generation of smart manufacturing platforms.
References
Farahani, M. A., Khan, M. I., & Wuest, T. (2025). Hybrid agentic AI and multi-agent systems in smart manufacturing. Preprint submitted to the 2026 NAMRC Conference. arXiv:2511.18258.
Xia, Y., Dittler, D., Jazdi, N., Chen, H., & Weyrich, M. (2024). LLM experiments with simulation: Large language model multi-agent system for simulation model parametrization in digital twins. IEEE International Conference on Emerging Technologies and Factory Automation (ETFA). DOI: 10.1109/ETFA61755.2024.10710900.