Hybrid Agentic AI and Multi-Agent Systems in Smart Manufacturing

This paper proposes and validates a hybrid framework that integrates LLM-driven agentic AI with traditional multi-agent systems to enable adaptive, interpretable, and cost-effective prescriptive maintenance in smart manufacturing through a layered architecture of strategic orchestration and specialized edge execution.

The Gist

Imagine a busy, high-tech factory floor. In the past, if a machine started acting up, a human expert had to rush over, check the gauges, remember past repairs, and decide what to do. It was reactive, slow, and relied heavily on human memory.

This paper proposes a new way to run that factory using a team of AI "robots" that work together like a well-oiled orchestra. The authors call this a Hybrid Agentic AI and Multi-Agent System.

Here is the breakdown in simple terms:

1. The Problem: Too Much Data, Too Many Decisions

Modern factories are covered in sensors (like digital eyes and ears) that scream data 24/7. Traditional computer programs are like rigid robots: they follow a strict rulebook. If the situation changes slightly (like a new type of vibration), the rulebook breaks, and the robot freezes. They can't "think" outside the box.

2. The Solution: A Team of Specialized AI Agents

Instead of one giant brain trying to do everything, the authors built a team of specialized AI agents. Think of this like a hospital emergency room:

• The "Brain" (LLM Orchestrator): This is the Chief of Surgery. It's a powerful AI (like a very smart, experienced doctor) that understands the big picture. It doesn't do the dirty work; it looks at the patient, figures out what's wrong, and tells the specialists what to do. It uses a Large Language Model (LLM) to "reason" and plan.
• The "Specialists" (SLM Agents): These are the nurses and technicians. They are smaller, faster, and cheaper AI models (Small Language Models) that live right on the factory floor (the "edge"). They handle specific tasks like cleaning the data, checking for missing numbers, or running quick calculations. They are fast and don't need to call the "Chief" for every tiny decision.
• The "Human Supervisor" (HITL): Just like in a hospital, a human doctor is always in the loop. The AI makes the recommendations, but a human has to give the final "Go" before any expensive or dangerous action is taken. This ensures safety and trust.

3. How They Work Together (The Workflow)

The paper tested this system on two real-world factory datasets. Here is how the team handled a crisis:

1. The "Eyes" (Perception Agent): The team looks at the raw data. It's like a nurse checking a patient's vitals. "Oh, the temperature is high, and the vibration is weird."
2. The "Cleaners" (Preprocessing Agent): Before the doctors can diagnose, the data needs to be cleaned. This agent fixes missing numbers and organizes the charts automatically.
3. The "Diagnosis" (Analysis Agent): The Chief Brain asks the specialists, "Which test should we run?" They try different mathematical models. If the first guess is wrong, the system doesn't panic; it tries a different model automatically until it finds the right answer.
4. The "Prescription" (Optimization Agent): Once the problem is found, the team doesn't just say "Machine X is broken." They say, "Machine X is critical. Send a technician now. It will cost $500 and take 2 hours. If we wait, it will cost $5,000."

4. Why This is a Big Deal

• It's Flexible: If the factory changes its layout or adds new machines, the AI team can figure out how to handle the new data without needing a human to rewrite the code. It's like a chef who can cook a new dish just by tasting the ingredients, rather than needing a new recipe book for every meal.
• It's Transparent: Old AI systems are "black boxes"—you put data in, and a result comes out, but you don't know why. This system keeps a "logbook" of every thought process. You can ask the AI, "Why did you suggest fixing Machine 4?" and it will explain its reasoning step-by-step.
• It's Fast and Safe: By using the "Chief" for big decisions and the "Nurses" for quick tasks, the system is both smart and fast. It keeps sensitive factory data local (on the factory floor) for security, only sending big-picture questions to the cloud.

The Bottom Line

This paper shows that we are moving from Predictive Maintenance (telling you a machine might break) to Prescriptive Maintenance (telling you exactly what to do to fix it, when to do it, and how much it will cost).

It's like upgrading from a car that just flashes a "Check Engine" light to a car that says, "Your tire is low, drive to the gas station on the left, and here is a coupon for a free rotation." The authors proved this "team of AI agents" works, is adaptable, and is ready to help factories run smarter, safer, and more efficiently.

Technical Summary

1. Problem Statement

The paper addresses the limitations of current maintenance and decision-making systems in Smart Manufacturing Systems (SMS) under Industry 4.0.

• Limitations of Traditional MAS: Traditional Multi-Agent Systems (MAS) rely on rigid, rule-based logic and static models. They lack the depth of reasoning, context awareness, and adaptability required for dynamic, unstructured industrial environments.
• Limitations of Standalone Agentic AI: While Large Language Model (LLM)-based Agentic AI offers superior reasoning and planning, single-agent architectures often struggle with scalability, latency, and the high computational cost of cloud-based inference. They also lack the distributed robustness needed for complex manufacturing workflows.
• The Gap: There is a lack of frameworks that effectively integrate the distributed coordination of MAS with the adaptive, high-level reasoning of Agentic AI to solve Prescriptive Maintenance (RxM) problems. Existing solutions often fail to bridge the gap between high-level strategic planning and low-level autonomous execution.

2. Methodology

The authors propose a Hybrid Agentic AI and Multi-Agent System (MAS) framework designed for RxM. The architecture is modular, layered, and combines LLMs for strategic orchestration with Small Language Models (SLMs) and rule-based agents for tactical execution.

A. Architecture Overview

The framework consists of five interconnected layers managed by specialized agents:

1. Intelligence Layer (LLM Orchestrator):
   • Role: Acts as the "brain," using an LLM (Google Gemini 2.5-Flash) for high-level reasoning, workflow state management, and tool selection.
   • Mechanism: It maintains context (goals, history, metrics) and generates structured instructions for downstream agents. It includes a self-correction mechanism (retry logic up to 3 times) and falls back to rule-based logic if the LLM fails.
2. Perception Layer:
   • Role: Data ingestion and schema discovery.
   • Function: Computes metadata (shape, types, missing values) and flags data quality issues to the Orchestrator.
3. Preprocessing Layer:
   • Role: Automated pipeline construction.
   • Function: Uses an SLM (e.g., Qwen3:4B via Ollama) and rule-based logic to perform schema discovery, feature analysis (correlation, mutual information), and dynamic selection of preprocessing techniques (imputation, scaling, encoding) based on data characteristics (e.g., memory size, missing value percentage).
4. Analytics Layer:
   • Role: Model selection and evaluation.
   • Function: Supports classification, regression, and anomaly detection. It employs an Adaptive Intelligence module: if initial model performance (e.g., $R^2 < 0.1$) is suboptimal, it automatically iterates through alternative model families and hyperparameters, logging all attempts for traceability.
5. Optimization Layer:
   • Role: Generating actionable recommendations.
   • Function: Translates model outputs into prioritized maintenance actions. It calculates priority scores based on statistical thresholds (Mean + 2/3 Std Dev) and feature importance, estimating costs and timeframes.

B. Workflow & Technology Stack

• Workflow: The system operates end-to-end: Data Loading $\rightarrow$ Schema Discovery $\rightarrow$ Dynamic Preprocessing $\rightarrow$ Model Training/Evaluation $\rightarrow$ Optimization $\rightarrow$ Human-in-the-Loop (HITL) Validation.
• Hybrid Execution:
  • LLM: Handles complex planning and context retention (Cloud-based).
  • SLM: Handles lightweight, low-latency tactical decisions locally (Edge-based), preserving privacy and reducing latency.
• Human-in-the-Loop (HITL): A CLI interface allows experts to review reasoning traces, approve/reject recommendations, and audit decisions, ensuring transparency and safety.

3. Key Contributions

1. Hybrid Architecture: A novel framework unifying MAS coordination principles with Agentic AI reasoning, enabling both distributed autonomy and strategic adaptability.
2. Task-Agnostic Intelligence: The system autonomously identifies problem types (classification, regression, anomaly detection) and adapts its pipeline without manual reconfiguration or task-specific scripting.
3. LLM/SLM Synergy: Demonstrates a practical trade-off where LLMs handle high-level orchestration while SLMs execute local, privacy-preserving tactical tasks.
4. Explainability & Auditability: Every agent decision is logged with a "reasoning trace," transforming opaque AI decisions into transparent, auditable processes for human operators.
5. Proof-of-Concept Implementation: A fully functional Python-based prototype with a GitHub repository, validated on real-world industrial datasets.

4. Results

The framework was validated using two datasets: the Smart Manufacturing Maintenance Dataset (SMMD) and the 6G-Enabled Intelligent Manufacturing Resource Dataset (6GMR). The system executed six workflows covering three problem types:

• Classification (SMMD): Predicted maintenance priority (High/Medium/Low).
  • Result: Achieved 97.2% accuracy. The system correctly identified "Downtime Cost" and "Vibration" as key predictors and generated 10 ranked, actionable recommendations.
  • Time: ~132.5 seconds.
• Regression (SMMD): Estimated failure probabilities.
  • Result: Achieved an $R^2$ of 0.9209 using a Linear Regression model (selected via adaptive intelligence). Generated prioritized actions for critical machines.
  • Time: ~165.3 seconds.
• Anomaly Detection (6GMR): Identified network-induced anomalies.
  • Result: Used Isolation Forest to flag 1,000 anomalies (1%) in 100k records. Correctly identified packet loss and latency as primary signals.
  • Time: ~127.4 seconds.

Key Findings:

• The system successfully generalized across different datasets and problem formulations without manual intervention.
• The Adaptive Intelligence module successfully triggered model retries when initial performance was low, demonstrating self-evaluation capabilities.
• The HITL interface successfully abstracted ML complexity, allowing non-experts to oversee the process.

5. Significance and Future Work

• Significance: This work bridges the gap between theoretical Agentic AI and practical industrial application. It proves that combining LLMs with distributed MAS can create systems that are robust, scalable, explainable, and adaptable to the dynamic nature of Industry 4.0. It moves maintenance from reactive/predictive to truly prescriptive and autonomous.
• Limitations: Current latency in LLM reasoning may hinder real-time streaming applications; validation was limited to offline datasets; and safety verification for unseen scenarios needs strengthening.
• Future Work:
  • Integration with real-time industrial protocols (MQTT, OPC UA) for streaming data.
  • Development of agent-to-agent communication for collaborative reasoning.
  • Expansion to multimodal data inputs.
  • Transitioning from generating recommendations to executing autonomous operational actions under human supervision.

In conclusion, the paper presents a foundational step toward autonomous, human-aligned manufacturing ecosystems, demonstrating that hybrid Agentic AI can effectively manage the complexity of modern smart manufacturing.