Maingen wants AI to survive a night shift at a solar farm
The Y Combinator startup builds reinforcement learning environments for industrial operations. Its first release, SolarBench, drops an AI agent into a simulated operations desk and asks a blunt question: can you keep the lights on and the money in?
The pitch for most AI agents is a chat window. You type, it answers, and if it gets something wrong you shrug and try again. Maingen, a two-person company in Y Combinator's Summer 2026 batch, is building for a place where shrugging is expensive. Its bet is that the next real test of an AI agent is not a clever prompt, but an unglamorous overnight shift running a fleet of solar farms.
The company builds reinforcement learning environments for industrial operations - the part of the economy that keeps power flowing, machines running, and orders shipping. That is a large and under-digitized slice of the country, roughly $5 trillion by Maingen's framing, and almost none of it exists as clean training data. You cannot scrape a factory floor. You cannot download the judgment of the engineer who decides, at 2am, whether a fault alarm is worth waking a technician for.
01 / THE PRODUCTSolarBench, or the exam nobody studied for
Maingen's first release is SolarBench, which the company describes as the first benchmark to test AI agents on long-horizon industrial work. The setup is specific. An agent is placed at a simulated remote-operations desk for a portfolio of solar sites. It gets an alerting system, per-site telemetry, work orders, standard operating procedures and manuals, an inventory, and an inbox full of owners and technicians. Then the shift begins.
The difficulty is not answering a question. It is figuring out what the question is. The task description does not tell the agent what to do. It has to read the noise, decide what actually matters, and act - all while keeping the portfolio profitable. Some of those actions cost real money in the simulation: sending a technician to a site or ordering a replacement part is priced, and it cannot be taken back.
That last part is what separates SolarBench from a trivia test. A multiple-choice benchmark rewards the right answer. SolarBench rewards a good shift - one where the agent triaged correctly, spent money only when it made sense, and left the portfolio better than it found it. The scenarios, Maingen says, come from hundreds of hours spent with the people who run solar for a living.
The messy physical world is the next horizon for agents. Maingen
02 / THE PROBLEMWhy the physical world resists AI
Language models learned to write by reading the internet. Industrial operations have no equivalent corpus. The knowledge is tacit, scattered across control systems, spreadsheets, phone calls, and habit. It is also unforgiving in a way text is not - a wrong call dispatches a truck, burns a budget line, or lets a fault sit until it becomes a failure.
For a reinforcement learning model, the bottleneck is not intelligence so much as a place to practice. You cannot let an experimental agent learn on a live power plant. Maingen's answer is to build the environment - a faithful enough simulation of the operations desk that a model can make thousands of decisions, get scored on the outcomes, and improve without anyone losing a megawatt. The environment, not the model, is the product.
03 / THE FOUNDERSTwo people, one operations desk
Maingen is led by Phillip Yan, chief executive, who previously worked at 0x, Scale AI, and Coinbase, and built exchange infrastructure that handled more than $150 million in volume. His co-founder and chief technical officer, David Yang, came through Scale AI and Amazon and contributed to mechanistic interpretability research cited by Anthropic for ICLR 2025. It is a pairing of someone who has shipped high-stakes financial systems with someone who has looked closely at how models actually reason.
The two-person size is not a footnote. It shapes the strategy: rather than build a broad platform, Maingen went deep on a single domain it could learn firsthand, sitting with solar operators and turning their decisions into a benchmark. Solar is the wedge, not the ceiling - the stated ambition is environments for industrial operations generally.
SolarBench measures whether a model can run an industrial operations desk. Maingen, on its first benchmark
04 / THE MARKETWhere Maingen sits
Maingen lands at an intersection that did not obviously exist a year ago: the RL-environment companies feeding frontier labs, and the operations software that runs energy assets. Against the first group, its edge is domain depth - real scenarios from real operators, priced and irreversible, rather than tidy puzzles. Against the second, its edge is that it is not selling a dashboard to human operators; it is building the training ground for the agents that might one day sit beside them.
The customers, then, are twofold. Frontier AI labs want environments and datasets to train and evaluate agentic models. Solar operators, independent power producers, and asset managers provide the ground truth that makes those environments credible - and stand to benefit if the agents get good. The business is B2B: sell the environment, the benchmark, and the data, not a consumer app.
05 / WHAT YOU CAN DO WITH ITFor labs, operators, and the curious
If you run an AI lab, SolarBench is a way to see whether your model can hold a shift instead of just answer a question, and Maingen's environments are a way to train toward that. If you run solar assets, the company is collecting the kind of operational reality that could eventually staff an on-call desk with software. And if you are simply watching where agents go next, SolarBench is a readable early marker: the frontier is drifting from the browser toward the boiler room. Maingen offers early access and dataset requests through its site.