Last day: 60 days building an AI agent for electrical engineering
Day 60 / 60
Ok. Last day.
On December 22, 2025, I wrote that I was going to build an agent capable of assisting in the preparation and review of electrical studies in 60 days. It sounded ambitious even to me. Many people believed an agent wasn't capable of doing real electrical studies — actual simulations, with engineering judgment, in professional software like PowerFactory. I believed it could. And after 60 days, I think I was right.
It's not perfect. But it works.
What D.N. achieved
The agent today can:
- Run power flows and short-circuit analysis autonomously
- Coordinate protections in small systems
- Navigate Infotécnica to extract real substation data from the SEN
- Query CNE and CEN regulations to make informed decisions
- Generate PDF reports of over 20 pages with tables, diagrams and analysis
- Run for over 60 minutes without human intervention
- Parallelize simulations across multiple machines running DIgSILENT in the cloud
I started with 4 tools on day 5. Today it has over 50. I started running on 1 virtual machine. Today it runs on 5 in parallel with a DAG that the planner itself builds.
About the studies: not all are equal
One thing I learned during these 60 days is that you can't talk about "electrical studies" as if they were all the same. The difference in complexity between them is enormous.
Power flows and short-circuits are studies the agent already does well. These are studies where the procedure is relatively clear: you configure the scenario, run the simulation, analyze the results, generate the report. There's judgment involved — which contingencies to test, how to interpret voltages, which scenarios are relevant — but the sequence of steps is predictable. The agent executes them autonomously, generates reports with results and the analyses are correct. I'd dare say these studies are practically solved for small systems.
Protection coordination is a different beast. Here the agent managed to coordinate relays in small systems, and that was the moment I felt I had something. But being honest, protection coordination has a complexity I haven't been able to fully tackle yet. It's not just running a simulation — it's understanding the topology, the protection philosophy, coordination margins, TCC curves, zone settings in distance relays, and making decisions that depend on each specific system's context. An experienced engineer makes those decisions with judgment accumulated over years. The agent is starting to do it, but it still has a way to go.
The gap between a power flow and protection coordination is the same gap between an agent following instructions and an agent having judgment. And that's the most interesting part of what's coming.
What I didn't achieve
I'll be honest: I haven't been able to test the agent on a real-scale case. I only have access to licenses for a maximum of 50 buses, and the SEN has over 2,500. Next week I should have access to a full license, and that will be the real test.
Protection coordination in large systems remains an open problem. The agent coordinates well in small systems, but I don't know how it will behave when combinatorial complexity explodes.
The moments that marked me
The day the agent coordinated relays on its own. Day 30. I gave it a complex prompt — run simulations, turn generators on and off, adjust relays. It ran for 9.6 minutes, used 37 tools, executed 46 steps. And it coordinated the relays that actually needed coordinating. That day I knew this was going in the right direction.
The day I lost an entire day because PowerFactory doesn't run on ARM. Day 5. Three Macs, none worked. I ended up on Google Cloud.
The first autonomous PDF report. The agent took 10 minutes, used 26 tools in sequence, ran 4 different contingency analyses. The report was mediocre, but it was real. Cost: $0.149 USD.
What building in public taught me
Over 500 people read this blog during the challenge. Double what my site got in an entire year.
But the most important thing wasn't the traffic. It was that documenting forced me to make progress. Every week I had to have something new to share. And that pressure, which I imposed on myself, was probably the reason I finished. There were weeks when I didn't feel like it, when the agent failed in absurd ways and I didn't know what to write. But I wrote anyway.
The other thing I didn't expect: building in public brought me interest from consulting firms that want to try and use what I'm building.
What I learned
About Chile's electrical system: Infotécnica, CEN, CNE, the open access processes... I understood it before, but now I understand it much better, like someone who had to code every rule and every exception.
About the studies: the real difficulty of each type of study, why protection coordination is so complex, why electrical engineers do what they do the way they do it.
About AI agents: that the number of tools matters. That Plan-and-Execute beats ReAct for long tasks. That an agent's memory is an incredibly hard problem. That this is the year of AI agents.
About myself: that I like writing more than I thought. That self-imposed deadlines work. That building something that solves a real problem in public, even imperfectly, feels different from anything else.
What's next
I'm taking the weekend off. On Monday, a new phase of the project begins.
Thanks to everyone who read and commented during 60 days.