LLMThings: Platform for Language Agents-Based Orchestration of IoT Systems

1Master Computer and Network Systems, Université Paris-Saclay
2Laboratoire IBISC, Université Paris-Saclay, Evry, France

LLMThings lets practitioners define crews of specialized agents, connect them to real IoT devices and tools, and orchestrate a live deployment from a natural-language intent.

Abstract

In this paper, we describe LLMThings, an experimentation platform dedicated to the study and evaluation of language agents-based orchestration of IoT systems.

The platform features a real-world IoT deployment while being modular, flexible, and interactive: it allows users to define (crews of) specialized agents, extend the system with various tools (e.g., devices, APIs, AI models, external services, etc.), run experiments, and monitor various metrics of the IoT system in real-time. This makes it a suitable testbed for education, research, and development purposes.

Motivating scenario: the smart nursing home

To illustrate how the distinct functional layers of an IoT orchestrator interact under real-world constraints, we present a concrete smart nursing home deployment. The physical infrastructure comprises a heterogeneous swarm of low-power wireless devices (in blue) operating on fragmented protocols such as Wi-Fi and ZigBee. The operational objective is driven by a high-level intent expressed by a nurse through the LLMThings interface:

“Make sure that the patient does not fall wherever she goes.”

As a patient (in red) moves from the “Activity Node” towards the garden, navigating unpredictable paths through the “Living” and “Kitchen” corridors with frequent stops, the orchestrator needs to coordinate services and resources in order to satisfy the nurse’s intent while complying with severe bandwidth, energy, and computational constraints.

Floor plan of a nursing home annotated with wireless devices, wireless transmissions, and two candidate paths of a moving patient.

IoT system monitoring patients within the nursing home.

Orchestration generally breaks down into:

  • service composition and workflow: decompose the initial intent into multiple independent tasks and dynamically link them into a single workflow;
  • resource allocation and placement: decide where to run computing tasks;
  • heterogeneity and interoperability: connect devices that use different languages;
  • mobility management, fault tolerance, and adaptation: maintain stable operations when devices move or when parts break or disconnect.

LLMThings

High-level architecture of LLMThings: a practitioner defines a crew of agents and the tools they can access; an end-user expresses an intent in natural language; the crew of agents draws on language models and MCP tools to orchestrate devices across the Cloud, Fog, Edge, and Devices layers.

Overview of LLMThings. Practitioners (e.g., sysadmin or developer) first define the agents, their coordination, and the tools they can access. The end-user (e.g., nurse) can then express an intent in natural language, which is processed by the resulting crew of agents to orchestrate the underlying IoT system.

LLMThings allows practitioners to study, design, and evaluate language agents-based orchestration approaches in a real-world IoT system. The framework provides a modular environment in which experimenters can define specialized agents, integrate heterogeneous tools, including IoT devices, AI models, and external services, configure realistic deployment scenarios, and monitor system behavior through quality of service metrics and interactive visualizations.

The platform is designed to facilitate rapid experimentation, reproducible evaluation, and educational demonstrations while remaining sufficiently flexible to accommodate diverse research questions.

The current deployment features a real-world IoT system composed of a dozen ESP32 devices interconnected via Wi-Fi and ZigBee and a set of approx. 30 sensors (e.g., camera, motion, and temperature).

Autonomic orchestration with MAPE-K

To evaluate LLMThings we investigate the use of crews of agents to implement various orchestration strategies and management loops. Notably, we implement the famous MAPE-K autonomic management loop composed of Monitor, Analyze, Plan, and Execute stages, as well as a shared Knowledge providing the information required by the different stages.

We define agents corresponding to each one of these stages and provide them access to specific tools, while enforcing a sequential coordination among them.

Acknowledgements

We want to express our gratitude to the department of computer science and the Direction de l’Offre de Formation (DOFREVA) of UEPS for their support.

BibTeX

@inproceedings{sadoudi2026llmthings,
  author    = {Sadoudi, Anais and Le Huyen-Trang and Hamidi Massinissa},
  title     = {{LLMThings}: Platform for Language Agents-Based Orchestration of {IoT} Systems},
  year      = {2026},
}