PowerAgent — open-source community

I founded the PowerAgent open-source community to develop shared infrastructure for agentic AI in power systems, combining fine-tuned foundation models, the Model Context Protocol, and benchmarks. It launched in April 2025 with support from the Power and AI Initiative (PAI) at Harvard SEAS.

1. 1industry problems shape benchmarks

2. 2benchmarks expose scientific gaps

3. 3open tools accelerate solutions

4. 4validated solutions create pathways to adoption

The ecosystem

In my role as the lead of the PowerAgent community, I cultivate the ecosystem to connect open research, workforce development, and industrial collaboration.

Open research

The benchmarks, PowerMCP and PowerSkills are developed in the open and published.

Workforce development

The PowerMCP and PowerSkills tutorial at the Harvard workshop on foundation models for the grid, and a guest lecture in ECE 285 at UC San Diego.

Industrial collaboration

Collaborators and users at ERCOT, ISO New England, MISO and EPRI; related work presented to Eversource, CenterPoint Energy and several national laboratories.

The PowerAgent roadmap was presented to the ERCOT Board of Directors, Technology and Security Committee, in December 2025 as A Roadmap Toward Agentic Intelligence in Power Systems, and PowerMCP and PowerSkills were taught as a tutorial at the 5th Workshop on Foundation Models of the Electric Grid (Harvard, March 2026) as Plug-and-Play: Connect PowerMCP and PowerSkills to Your LLM. The roadmap paper (IEEE Power & Energy Magazine, 2025) was also selected as one of the magazine’s most popular articles.

I view LLM-based agents as an orchestration layer that connects validated models, data, and engineering tools without replacing physical models or human authority. Our current work makes PowerAgent explicitly grid-native by embedding grid topology, knowledge graphs, and unified grid data schemas, creating intelligence specialized for grid engineering rather than generic agents. The platform is being developed with ISO New England’s advanced technology team and tested by transmission-planning engineers.

To evaluate whether these agents can perform reliable engineering work, we also develop physics-grounded benchmarks, including PowerAgentBench-SS, PowerAgentBench-Dyn and RestoreBench, to test contingency analysis, dynamic studies, and non-convergent power-flow restoration through independent physical validation. All three are in the publication list, and the research behind the platform is thrust III, Deployable Agentic Intelligence.

Beyond PowerAgent, I am a core contributor to three open-source projects for power system modeling and analysis.