China's grids use AI agents for dispatch but keep humans in loop

China runs the largest and most complex electricity system on Earth, and it is quietly moving the hands off the controls. For decades, keeping the lights on meant veteran dispatchers reading screens, judging risk from experience, and issuing switching orders one at a time. That model is now being supplemented, and in narrow tasks replaced, by software agents that can read the whole grid, propose actions, and in some cases carry them out, with the human reduced to a reviewer rather than the operator.

The pressure behind the change is structural. Wind and solar now make up a large and growing share of generation, while storage, charging piles, and millions of flexible loads have turned a one-way delivery network into a multiplayer, real-time game. A skilled dispatcher can weigh perhaps ten variables at once; a modern scheduling decision can depend on thirty-five or more. When a heatwave pushes load up and renewable output swings minute to minute, no human can re-optimize the network fast enough. That is exactly the gap large language models and agent frameworks are being pointed at.

State Grid's flagship is the Guangming ("Light") power large model, described as the country's first industry-scale, hundred-billion-parameter multimodal model built for electricity. In day-to-day use it can produce several load-transfer plans within seconds and recommend the best one; operators reportedly adopt its suggestions more than 90% of the time. Wired into multi-agent coordination across provincial, prefecture, and county dispatch, the system has compressed day-ahead verification from several days to under two hours and cut manual workload by roughly three quarters. A dispatch decision agent trialed at the national control center handles alarms with about 95% accuracy and shrinks decision time from at least fifteen minutes to minutes. In Changsha, an "explainable" dispatcher built on the same base model states not just what to do but why, flagging risks, affected customers, and alternatives, a design choice meant to keep the human able to audit every move.

China Southern Power Grid has pushed the same idea into planning. Its "Da Watt Yunrui" distribution-network planning agent, fully deployed in Yunnan, has run more than ten thousand risk diagnoses and can turn out a planning report in about ten minutes instead of days, while lifting planning efficiency by roughly 80% and trimming the overload rate by around a fifth. The product was named one of the ten "treasure of the exhibition" picks at WAIC 2026 and also took an award at the AI for Good Global Summit in Geneva. A separate "Xunjie" agent digests tens of thousands of data rows in ten seconds during the summer peak.

None of this is limited to the two giants. A trading agent from GuoNeng RiXin, branded "Kuangming," is live across nineteen provinces, helping generators and storage players bid into electricity markets. Langxin's "Jiugong" trading agent forecasts prices with better than 90% accuracy. In Henan, a virtual-power-plant platform called "Great White Shark 2.0" runs on thirty-five coordinated agents around the clock with no human on shift, hitting 95% load-forecast accuracy across more than two hundred forty connected resources. Xiong'an's main-and-distribution dispatch agent has shown second-level fault self-healing since its February 2026 trial, and Jiangsu's Taizhou claims the country's first conversational autonomous dispatch agent for grid drones. The pattern is spreading from showcase to routine work.

Policy has cleared the lane. In April 2026, four ministries, the National Development and Reform Commission, the National Energy Administration, the Ministry of Industry and Information Technology, and the National Data Administration, issued an action plan on bidirectional AI-energy empowerment, carrying out the State Council's broader "AI+" directive. It sets 2030 goals, opens high-value energy scenarios to AI, and calls for unified data classification and sharing standards.

The hard ceiling is not accuracy. It is trust and structure. Grid data sits in silos owned by different market players, generators, dispatchers, and consumers, with no common standard, so no single agent can see the whole picture. More decisively, there is no established mechanism to trace or assign blame when a fully autonomous decision goes wrong. That hesitancy is visible in how the systems are built: Guizhou Power Grid, for example, grants its agents only tiered rights, to read, to analyze, to act, and keeps a "brake" that can halt a high-risk move and hand it back to a person. A model that predicted load at 98.7% accuracy in one province reportedly stumbles elsewhere, because every regional grid "speaks its own language." Autonomous assistance, not autonomous authority, is the deliberate design.

[Analysis] Why is certifying autonomous grid dispatch harder than certifying a self-driving car or a factory controller? Three reasons. First, scope of failure: a car that errs can pull over; a dispatch error propagates across millions of nodes within seconds and can cascade into a regional blackout, so there is no contained test track. Second, ownership: a vehicle has one manufacturer to certify and insure, but the grid's decision touches generation, dispatch, and load entities that are separate commercial bodies, so no single principal can sign off on, or be liable for, an autonomous action. Third, the environment is far less bounded than a road or a process line: weather, market prices, and distributed generation shift constantly, and a model trained in one region often fails to generalize, making scenario-based certification unreliable. Until cross-entity liability frameworks and traceability standards exist, "human in the loop" is not a stopgap but the certification itself.