-Title: Placement of LLM orchestrators in holonic multi-agent AI–digital-twin systems: a comparative study across manufacturing workloads
-Journal/Conference: 
-Authors: OUNGSUB KIM, Jongpil Jeong
-DOI: 
-Journal/Conference Link: 

Abstract: Large language model (LLM) agents are increasingly proposed as orchestrators for multi-agent, digital-twin-enabled manufacturing, but the question of where such an orchestrator should be placed within an existing holonic control architecture—and under which workload conditions a given placement is advantageous—has not been addressed. This paper formalizes the problem as the insertion of a new "L-holon" into a PROSA holarchy and derives a nine-cell compatibility map—three collaboration paradigms crossed with three holonic placement layers—that yields three high-viability configurations: hierarchical command–control at the staff layer (A×Staff), conversational peer orchestration at the order layer (B×Order), and a role-based collaborative pipeline at the staff layer (C×Staff). A pre-specified measurement contract fixes six comparison dimensions and the exact simulator outputs required to compute them, so that the empirical study is defined before implementation rather than dictated by whatever the simulator happens to emit.
Using a PROSA-inspired discrete-event simulator swept across eight workload cells (disturbance × scale × complexity) with thirty paired seeds per cell under uniform fairness caps, and with per-decision LLM interaction costs measured from a real local model (Qwen2.5 via Ollama) and injected into a calibrated model, the three configurations are compared against holonic and non-holonic baselines under a pre-registered non-parametric analysis. No single placement dominates: the best configuration crosses over on task complexity—C×Staff minimizes weighted tardiness on simple workloads while A×Staff wins on complex workloads, independent of disturbance and scale (Kruskal–Wallis p ≈ 10⁻²¹ in every cell)—while conversational peer orchestration (B×Order) incurs the highest communication overhead at roughly 2,710 tokens per completed order. The optimal placement of an LLM orchestrator in a holarchy is therefore not absolute but workload-conditional, and the compatibility map together with the measurement contract provides a reusable basis for making the placement decision.

-Status: Submitted (2026/07/02)