For the past years, predictive maintenance has become essential in mitigating unplanned downtime in smart factories. Predictive maintenance solutions detect early threats in equipment before they become critical. Although several traditional predictive maintenance methods exist, inspired by artificial intelligence (AI) and machine learning (ML) techniques, they often rely on singular and opaque machine learning models, lacking flexibility, scalability, and adaptability to real-time manufacturing demands. This study proposes a Multi-Agent Predictive Maintenance System that integrates three intelligent agents—Data Analytics, Diagnostics, and Scheduler—to deliver a robust (via modular and high-performing ML components), interpretable (by identifying root causes through feature attribution), and operationally efficient (through risk-prioritized scheduling) solution. We create a Data Analytics Agent to preprocess real-time sensor data from the publicly available UCI AI4I 2020 Predictive Maintenance dataset and build a Random Forest binary classifier that effectively detects machine failures. The model achieved 99.8% accuracy and high precision-recall performance, ensuring dependable fault prediction. To support maintenance decision-making and transparency for the system, we also identify the root cause of machine failures, such as torque, process temperature, and tool wear, by performing feature attribution analysis on the trained random forest model via a diagnostic agent. We develop a Scheduler Agent to prioritize maintenance tasks through a risk-based queue, scheduling the top ten highest-risk machines for service using predicted failure probabilities. The results demonstrate the potential of multi-agent AI systems in transforming traditional predictive maintenance systems into intelligent, proactive, and fully automated processes in innovative manufacturing environments.