A grid-operation agent that talks you through saving the power grid.
An LLM agent that returns an overloaded power grid to a safe state while narrating its
reasoning — built on Grid2Op with a 118-bus demo scenario and a
multi-agent advisor ring (Weather, Asset Health, Screening, Grid Events). It observes the
grid, searches remedial actions, simulates them with a real power-flow solver, consults its
advisors, and applies only protocol-approved actions.
The cockpitThe operator view: a 118-bus transmission grid drawn as a live blueprint, every substation, line and generator wired up and humming. Beside it runs the agent comms panel — a transcript that updates in real time as the agent reads the grid and starts to think. Right now everything is calm and balanced. The interesting question is what happens when it isn't.The crisisA contingency hits — demand spikes, a line trips — and power reroutes onto neighbouring lines that were never sized for it. They flag red as they cross their thermal limits. In a real grid this is how blackouts begin: one overload trips a protection relay, which dumps even more load onto the next line, which trips in turn — a cascade that can darken a region in seconds. The agent now has only a short window of simulated time to find a move that pulls the network back from the edge before the dominoes start to fall.The rescueThe agent doesn't guess. It tests candidate moves against a real power-flow solver, runs the survivors past its advisor ring, and applies only what clears the operating protocol. The overloads release, every line drops back under its limit, and the grid settles. And because each step was argued out loud, the rescue isn't a black box that flipped some switches — it's a decision an operator can read back, line by line, and trust. That is the whole point.