The Pulse
Grid Studio’s Mean Benchmark-Family Score Was 99.9%
Grid Studio scored highest on a suite of 519 power-system tasks drawn from eight published benchmark families, a paper reports. Its 99.9% figure is the mean of the benchmark-family scores, not a pooled score across all tasks.

Grid Studio scored highest on a suite of 519 power-system tasks drawn from eight published benchmark families, according to a paper on Grid Agents. Its reported 99.9% is the mean of the benchmark-family scores, not a pooled score across all 519 tasks.
Source arxiv.org
Grid Agents has two components. Grid Ontology is a searchable knowledge system for regulatory, standards and engineering material, with answers grounded in cited evidence. Grid Studio is an agent system for power-system engineering. It chooses studies and reads their results, while established solvers compute the numbers it reports. The authors describe the system as one in which engineers retain decision authority.
The power-system suite combines PowerAgentBench-SS, PSAB, PGLib-OPF, PowerCodeBench, PFBench, RL2Grid, PFΔ and ProOPF-B. The tasks cover networks ranging from 4 to 9,241 buses. They vary in the work required: some run a study on a published network, while others apply changes, identify what to change, rebuild a published operating point or modify an optimisation problem from a written request. None of the 519 tasks requires a multi-stage study, which the paper evaluates separately.
The authors use three elements to score engineering results: numeric closeness, answer accuracy and a strict pass rate. A task counts as a strict pass only when the answer is within the benchmark’s tolerance and the recorded work shows that the required study ran on the required grid state. The verification checks vary by system. For Grid Studio, the paper replays tool calls to check whether the required study ran on the required grid state; that is not described as the general method for every system in the comparison.
Grid Studio was also the fastest system compared, with a statistically significant difference, the paper reports. That timing comparison has a qualification: it excludes Grid Studio’s solver startup while counting some startup work for competing agents. The speed result therefore comes with an asymmetry in what startup work was included.
The Grid Ontology holds legislation, grid codes, standards and process documents. The Grid Ontology also has memory for facts, skills and experiences learned through use, with changes affecting other people taking effect only after approval. In its regulatory evaluations, the system cites relevant evidence more often than a general coding agent or retrieval pipelines. Results are also reported on LegalBench-RAG, the Enterprise RAG Challenge and ViDoRe V3.
Grid Studio focuses on power-system studies, including power flow, outage checks and optimisation. The benchmark suite brings together previously separate published families and grades tasks on a shared scale based on the work needed to reach an answer. The study’s broader aim is to assess not just whether an answer is correct, but whether it is supported by the work performed on the appropriate network state.
The authors write, “In both, engineers make the decisions.” With an optimiser in the loop, reinforcement packages cost 33–67% less on average than packages designed by the agent alone, the paper reports.
Sources
- Original source · Preprint arXiv arXiv:2611.03534