Multi-Agent Large Language Models for Distributed Internet of Things Analytics in Smart Environments
Aygün Varol, Asif Shaikh, Naser Hossein Motlagh, Mirka Leino and Johanna Virkki · SpliTech 2026, IEEE
A team of small language-model agents on cheap edge computers answered questions about a building's sensors perfectly until one agent was knocked out or turned hostile, which cut accuracy by more than a third.
Abstract
Internet of Things (IoT) sensors in smart environments produce large-scale data streams that can enable intelligent automation and user-centric services. However, utilizing Large Language Models (LLMs) for IoT data processing in resource-constrained indoor settings raises challenges in balancing computational efficiency, response latency, and service quality across heterogeneous computing resources. In this paper, we present a hierarchical multi-agent architecture in which a supervisor agent running on an edge device delegates tasks to location-specific worker agents deployed on Raspberry Pi devices. The agents collaborate to distribute analytics workloads across heterogeneous hardware, while Retrieval-Augmented Generation (RAG) grounds LLM outputs in the underlying sensor measurements. We deploy BME680 sensors in three locations (office, kitchen, hallway) and demonstrate feasibility through an experimental deployment at Tampere University. Experiments across 135 queries show that the system achieves 100% accuracy under normal operation, retains 88.9% accuracy under degraded operation with one agent offline, and drops to 62.2% under adversarial interference within the multi-agent system.