Skip to content
Open access

Hierarchical Multi-Agent LLM Architecture for IoT Platform Configuration: Design Patterns and Production Lessons

Nov 2026 · IoT · 0 citations · 28 references

Abstract

Configuring an Internet of Things (IoT) platform produces interdependent device definitions, relationships, access controls, data processing rules, alerting rules, and dashboards. Commercial platforms generate such artifacts from natural language, but published evidence on their performance in use is limited. We analyze five months of production logs from a deployed system with 84 schema-constrained agents organized in three levels. Agents declared missing dependencies through a backfill channel in 8.7% of rule generations and 55% of propagation rule generations. The system completed 1408 configurations for external users. Of the 1376 with an installation attempt, 768 (55.8%) had no recorded failure at any stage. Of the 1374 with attempt-level records, 791 (57.6%) had no recorded failure on their first installation attempt. The dependency rule attributed detected faults to the generator in 10.2% of completed configurations (95% CI 8.7–11.8%). An audit of a stratified sample labeled independently by two authors yielded estimates of 11.5% and 12.6% (κ=0.87). Detected faults were concentrated in data processing rules. Per-configuration medians were 1012 s of cumulative agent processing, 347 s from session start, and $0.57 of inference.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.