[AI SHAPERS 2025] Autonomous Machines and Multi-Agent Systems: Towards Intelligent Maintenance - Loïck Jeanneret

Today, modern industrial machines can already detect and resolve certain anomalies thanks to their integrated sensors. For example, when a temperature sensor detects an abnormal value, the machine can adjust the fan speed to correct the problem. If this is not enough, an alert is sent to a technician.

However, these actions remain limited to pre-programmed scenarios. Resolving complex faults still largely depends on human intervention, which requires in-depth knowledge of the machines.

In the near future, if artificial intelligence continues to advance at its current rate, it will become possible to design even more autonomous systems. By combining sensors, AI, and multi-agent architectures, these systems could achieve a level of proactive and autonomous maintenance superior to what we know today.

Autonomous problem solving using a multi-agent system.

A multi-agent architecture consists of several specialized agents capable of working in parallel. Each agent is responsible for a specific aspect of problem diagnosis and resolution, and has tools at its disposal to obtain additional information. These sub-agents are coordinated by an orchestrator agent, which is responsible for analyzing the problem, establishing a diagnosis, and managing its resolution.

The number and type of sub-agents vary depending on the machine and maintenance objectives. However, several generic agents can be considered.

Example of an architecture for solving a problem:

A monitoring module continuously analyzes sensor data. In case of an anomaly, it alerts the orchestrating agent, transmitting the relevant information to initiate diagnosis. The orchestrating agent then begins collecting further information.

To achieve this, he uses several sub-agents in parallel:

• An agent consults the technical manual to identify the relevant sections related to the problem, then synthesizes the useful information without having to analyze the entire document.

• A second agent examines the machine's maintenance history to check if a similar failure has already occurred and how it was resolved.

• Finally, an agent with general knowledge suggests common causes that could explain the anomaly.

Once the sub-agents have completed their analyses, the orchestrator synthesizes the results and coordinates the generation of action plans. These plans consist of actions that the machine can execute, and, if necessary, tasks requiring human intervention.

The orchestrator then evaluates the different action plans and selects the most suitable one. It then requests an agent to execute it. If certain steps cannot be performed automatically, an agent triggers an alert clearly explaining the actions to be taken by an operator, while also giving them the opportunity to ask further questions.

Once the problem is solved, an agent is responsible for documenting and recording all the steps taken, thus enriching the machine's technical memory.

Technological challenges and limitations to overcome

Several technical, human and organizational obstacles still need to be overcome before this scenario can materialize in an industrial environment.

In my opinion, generative artificial intelligence should serve humanity without replacing or endangering it. It is valuable when it increases productivity in controlled contexts: when the generated information is verifiable and verified, or when the consequences of an error remain limited.

For example, as a developer, I regularly use a language modeling tool to complete code. However, I always make sure that each generated line matches what I want to produce. Frequently, the generated code, while seemingly coherent, is incorrect, contains security vulnerabilities, or is difficult to maintain.

In an industrial environment, where an error can lead to material damage or injury to people, it is not yet conceivable, in my opinion, to let an AI act without a minimum of supervision and limitations.

Furthermore, these environments often present technical constraints. For example, industrial machines are rarely connected to the Internet, which can prevent AI from retrieving information online.

However, solutions do exist. Regarding security issues, restricting AI to only perform predefined and verifiable actions reduces risks, although this limits its scope of action.

Regarding connectivity constraints, a centralized server internal to the company could provide significant computing power at a lower cost, while also offering an internal knowledge database, thus ensuring compliance with confidentiality requirements.

The future of industry will undoubtedly be more autonomous and smarter. But it remains essential to keep current limitations in mind, while remaining attentive to upcoming technological developments.

An article by Loïck Jeanneret, taken from the collective book "Bots and Robots", as part of the AI Shapers 2025 selection.