The Pulse
Grid Studio Scores 99.9% Across 519 Power-System Tasks
A paper submitted to arXiv on November 2 describes Grid Agents, a system combining regulatory research with power-system engineering studies. Its Grid Studio agent scored 99.9% across 519 benchmark tasks, while the authors note that real pl
AI.info Team · Written with AI · fact-checked ·
Grid Studio scored 99.9% across 519 power-system tasks, but its authors also describe a gap between benchmark networks and the larger planning models used by operators. The system, part of Grid Agents, ran those tests faster than the alternatives in the study, while engineers retain decision authority. MaoXun Huang, Kshitij Kapoor and Eli Louise submitted the paper to arXiv on November 2, 2026.
Grid Studio pairs agents with established solvers
The paper presents Grid Agents as two linked components: Grid Ontology, which stores searchable regulatory, standards and engineering knowledge, and Grid Studio, which runs power-system studies in a dedicated engineering environment. The agent selects and sequences studies, while established solvers calculate the reported numbers. Answers drawn from Grid Ontology cite supporting evidence, and the system’s memory can change as agents learn from use; changes affecting other users require human approval.
The authors frame the division of responsibility plainly: “Agents prepare the work and people decide.” Their evaluation also checks more than final answers: it replays tool calls to verify that a system ran the required study on the required grid state.
Its 519-task lead includes a speed advantage
The researchers assembled eight published power-system benchmark families into a 519-task suite, spanning networks from 4 to 9,241 buses. Grid Studio achieved a 99.9% end-to-end score, with 100% on seven benchmark families and 99.2% on PowerCodeBench. The paper says the result was statistically significantly higher than every other system compared.
Grid Studio’s median agent work time was 13.9 seconds per task. Codex CLI scored 99.4% and took a median 21.6 seconds; Power-Agent also scored 99.4%, with a 16.8-second median. The paper’s end-to-end measure combines numeric closeness, answer accuracy and a strict pass that requires both a sufficiently accurate answer and evidence that the right study ran on the right network state.
Regulatory retrieval wins, but by narrow margins on some boards
Grid Ontology was evaluated separately on regulatory retrieval benchmarks. On LegalBench-RAG, it achieved an F1 score of 0.980, ahead of the baselines in the paper. On the Enterprise RAG Challenge, it scored 124.5 out of 150, compared with 123.7 for the best published entry; on ViDoRe V3, its score was 70.23 against 70.01 for the best published pipeline.
For the study’s own GB Regulation Set, the paper reports evidence-retrieval recall of 0.967 for Grid Ontology and 0.856 for Claude Code. Both systems answered every question correctly there, the authors write, so the comparison focused on whether they cited the right supporting evidence.
The authors flag a scale limit
The strongest headline result applies to the paper’s benchmark suite, not to a full connection assessment on an operator’s largest planning model. The tested networks reached 9,241 buses, while the authors say an operator’s planning model can run to tens of thousands, and a connection study can chain multiple analyses. On the 109 benchmark tasks involving networks above 2,000 buses, Grid Studio passed 108; Power-Agent passed 106, Codex CLI 90 and Claude Code 74.
The team separately tested reinforcement planning in five studies on networks of 30, 118 and 2,000 buses. With an optimiser in the loop, the resulting reinforcement packages cost 33–67% less on average than packages designed by the agent alone, according to the paper. The benchmark scores and cost comparison are the paper’s reported results; engineers, the authors say, make the decisions.