Harmonizing Real-Time Constraints and Long-Horizon Reasoning: An Asynchronous Agentic Framework for Dynamic Scheduling

Researchers introduce RACE-Sched, an asynchronous AI framework that combines real-time symbolic heuristics with LLM-powered reasoning to solve dynamic job shop scheduling problems in industrial systems. The approach decouples fast reactive execution from slower deliberative optimization, enabling superior performance over deep reinforcement learning baselines while maintaining interpretability and millisecond-level response times.

RACE-Sched addresses a fundamental tension in industrial automation: the need for instantaneous decision-making in control systems conflicts with the inference latency of advanced AI models. Traditional priority rules lack flexibility for complex disruptions, while learning-based systems often sacrifice explainability or fail to generalize across different problem configurations. This research demonstrates a pragmatic solution through architectural separation—a reactive stream handles immediate dispatching using fast symbolic heuristics, while a parallel deliberative stream leverages LLMs to continuously improve these rules without blocking operations.

The framework reflects broader trends in AI systems engineering where practitioners recognize that single-model approaches cannot simultaneously optimize for speed, interpretability, and performance. By using LLMs as offline rule synthesizers rather than online decision-makers, RACE-Sched achieves the reasoning benefits of large models without imposing latency penalties. The semantic rule repository adds transferability across problem scales, addressing a known limitation of many deep learning scheduling approaches that struggle with generalization.

For industrial automation stakeholders, this work suggests that hybrid symbolic-neural architectures may outperform pure learning methods for time-critical optimization tasks. The approach maintains human-interpretable decision rules—crucial for regulatory compliance and operator trust in manufacturing environments. Enterprises evaluating AI scheduling solutions should consider whether this asynchronous pattern could apply to their control architectures. The validation across multiple benchmarks (GEN-Bench, MK-Bench, JMS-Bench) indicates maturity beyond theoretical concepts, suggesting near-term applicability in manufacturing operations.

• →Asynchronous dual-stream architecture decouples real-time execution from deliberative reasoning, enabling LLM-powered optimization without latency penalties
• →Framework outperforms deep reinforcement learning and LLM baselines on multiple scheduling benchmarks while maintaining interpretable symbolic rules
• →Semantic rule repository enhances transferability across different problem scales and complexities
• →Atomic rule deployment ensures safety and prevents disruption to control loops during continuous optimization
• →Approach balances the speed requirements of industrial control systems with advanced reasoning capabilities needed for complex dynamic events