Large language model multiagent systems for industrial problem discovery and prioritization

A persistent gap separates academic research priorities from the operational challenges faced by manufacturing firms. Established identification methods, among them Failure Mode and Effects Analysis (FMEA) workshops, Lean audits and consultant engagements, are episodic, bounded by participant expertise and organised along departmental lines, which makes cross-departmental issues hard to surface. Multi-agent systems built on large language models (LLMs) have meanwhile shown collaborative problem-solving ability in software engineering and, more recently, in shopfloor scheduling and supply chain coordination. This article is a Perspective. It develops a framework and a research agenda; implementation and empirical evaluation fall outside its scope and are set out as its first research priorities. Three contributions are offered. The first is a synthesis of literature from industrial problem identification, LLM-based autonomous agents and manufacturing AI, which positions the proposal against existing industrial LLM agent work. The second is a framework in which six domain-specialised agents (production, quality, maintenance, supply chain, safety, energy) would collaborate through a five-phase protocol: worker-driven discovery, cross-departmental analysis, candidate-relationship detection, consensus-based financial assessment and debate-driven prioritisation. The prioritisation procedure is fully specified, including the normalisation of financial estimates, a measure of estimate dispersion, the derivation of criteria weights, a non-compensatory gate for safety-critical items and a mandatory human verification step. The third is a worked calculation on a hypothetical steel manufacturing vignette, included so that the procedure can be reproduced and scrutinised rather than as evidence of effectiveness. A preliminary qualitative readiness assessment and a three-horizon agenda of fourteen research questions complete the paper. The benefits identified are presented as hypotheses for that agenda to test rather than as demonstrated outcomes.