France’s national grid operator, RTE (Réseau de Transport de l’Electricité), has stopped playing around with chatbots and moved on to serious business. According to a report by Jérôme Picot and Clément Goubet, the company is deploying agentic systems to manage critical physical infrastructure. This isn't about text generation; it's about the direct integration of LLMs with industrial simulators via the Model Context Protocol (MCP). This shift transforms neural networks from mere advisors into executors capable of handling complex engineering calculations.
Tech stack: pypowsybl-mcp
At the heart of the architecture lies pypowsybl-mcp. This interface "tunnels" the capabilities of the pypowsybl power grid modeling framework directly to AI agents. RTE experts estimate that manual data preparation and scenario configuration have become an unaffordable luxury in the face of surging renewable energy—engineers simply cannot keep up with the volume of hypotheses required.
Use of multi-agent workflows where roles are clearly divided between planners and executors. Automation of simulation setup and result extraction. Maintenance of full oversight and control by human engineers.
Paradigm shift: Action Models over dialogues
This represents a major pivot from recreational dialogues to Action Models. Instead of trying to teach a neural network physics, it is given a standardized tool to interact with professional software.
This approach addresses security and scalability concerns: agents operate within a closed-loop engineering environment under human supervision, fully complying with EU AI Act requirements.
Results and outlook
Instead of trying to boil the ocean, French engineers have built a specialized translation layer. This effectively turns an LLM into an operating system for power grids, where various tools communicate through a unified format. In our view, it is this kind of "grounded" routine automation—rather than mythical breakthroughs—that will form the real foundation for managing complex state systems in the coming years.