Large language models (LLMs) are pushing agents beyond fixed workflows, enabling autonomous planning and execution and opening new opportunities for smart manufacturing. Yet the dynamic shop floor, intricate physical constraints, and strict reliability demands make direct deployment of generic agents challenging. This paper offers a systematic overview of LLM‑driven industrial agents. It first identifies five key challenges: real‑time perception, physical constraint modeling, reliability assurance, cross‑system coordination, and safety control. Then it defines the notion of an industrial agent, highlighting essential traits such as embed‑ability, interpretability, closed‑loop control, and deep integration with legacy manufacturing systems.
A unified taxonomy is introduced, covering three application scenarios (production scheduling, equipment maintenance, quality inspection), two embodiment forms (software‑only service, soft‑hardware hybrid), and five capability levels: perception, understanding, planning, execution, and learning. Each level maps to concrete techniques—multimodal data parsing, semantic extraction, LLM‑based task decomposition, robot/PLC command issuance, and feedback‑driven optimization.
Based on this taxonomy, a four‑layer system architecture is proposed:
- Perception Layer – gathers and preprocesses multimodal sensor data into a unified state vector;
- Cognition Layer – leverages LLMs to translate natural‑language intents into executable task graphs;
- Execution Layer – maps task graphs to concrete machine commands or workflow steps, with a scheduler handling parallel control;
- Monitoring Layer – continuously collects execution feedback, performs anomaly detection, and triggers closed‑loop corrections. Core modules (state synchronizer, task orchestrator, constraint solver, adaptive learner) create a full perception‑to‑learning loop.
The survey then reviews representative applications across the industrial lifecycle: in production planning, LLMs generate updated schedules in response to order changes; in equipment maintenance, agents infer root causes from fault logs and produce repair instructions; in quality inspection, vision models combined with LLMs automatically generate defect descriptions. Finally, future research directions are outlined, emphasizing high‑reliability verification, cross‑domain knowledge transfer, and compliance‑oriented safety certification.
Review: By linking challenge analysis to concrete system design, the article offers a structured reference for developing LLM‑based industrial agents, providing a solid foundation for subsequent research and implementation.