The 4 Levels of Agentic AI for Power Systems

General
1

Level 1 - Non-agentic AI

Use Copilot, Gemini, ChatGPT, or others to ask questions about how to run transmission planning studies. Read and interpret the output of the LLM and apply it to the problem at hand. It is not able to take any action on your behalf and has no access to tools.

image
image841×585 111 KB

Level 2 - Agentic Chat Interface

At IEEE GM in Austin last July, I was introduced to PowerAgent which as far as I’m aware is the first agentic AI power systems tool. The most basic level is a chat interface through mcphost, Claude Desktop, etc. This allows the user to describe a task to the agent and wait for a reply, for example asking it to open and solve an IEEE 39 bus system.

image
image1453×1005 193 KB

Level 3 - Agentic Tool Calling from Defined Process

The next iteration was manual workflow creation through tools like LangGraph or especially n8n. I didn’t learn of these until around 6 months ago when I began working on PowerWF from the same authors as PowerMCP. This allows a PowerSystems engineer to program a process, like real time contingency analysis, to be run on a set trigger or schedule, for example fetch a case from an FTP server and run contingency analysis every 5 minutes then summarize the results to the operator.

image
image1507×558 102 KB

These tools support features called “human-in-the-loop” where a human can be asked before every tool call.

Level 4 - Fully Autonomous Agents

The latest iteration of agents is fully autonomous through tools like OpenClaw. This has gone especially viral around a month ago but the tools have been around for several months already. These can combine numerous tool calls, do basic problem solving, remember long context windows, and more. It will eventually be possible to give a basecase and ask for a certain type of study to be performed and get a draft study report in mere minutes. Or request a model for a piece of equipment and allow the agent to search the internet for datasheets and generate a model that replicates the behavior in the powerflow tool of choice.

image
image1789×639 66.5 KB

image
image621×629 74.5 KB

Discussion

What does everyone think about the application of tools like these?

Is there anything exciting about the state of the tools today?

Where do we see the tools going in the future?

Is there a Level 5 automation tool out there?

What are some ways the tools can be improved?

What redlines should be encoded into fully autonomous agents like OpenClaw?

2

That’s a great summary!

Let’s imagine 5-10 years, if every power engineer has a general LLM on their laptop (like claude code), what’s the fastest way to make it power engineering-useful? Maybe MCP and Agent Skills are the only 2 plug-in they needed.

3

I agree if this tech was accessible to the point that power engineers can run powerful, local, agentic AI that could remove a lot of the modeling, scripting and simulation friction that's probably all we need for the next 5 years. But the tools need to get packaged into a product that people can install easily. It can't be a 10 page long tutorial in the command line like it is right now. That's what AWS Hadron is doing right for sure.