Epsilab launches from YC to build RL environments for enterprise work

A frontier AI lab trains an agent on a benchmark. The benchmark scores look good. The lab releases a model. A user sits down at a real office task. The agent fails inside a minute. Epsilab launched on Y Combinator today with a bet on where that gap comes from. The company says the fault is the training ground. It builds reinforcement learning environments from real-life data and the work of named industry experts. Its customers are the labs. Its YC launch had 41 votes at the time of this item.

Epsilab is in the YC Fall 2026 batch. The founders listed on the launch page are Joanne Wang, Fengjiao, Evan, and Matthew. Joanne and Fengjiao are second-time YC founders. Fengjiao is a reinforcement learning researcher. The page says she worked at the MIT Media Lab and at Vimeo. Evan is a former CTO of another startup. He did LLM research at UC Berkeley. Matthew worked as a quant infrastructure engineer at Tower Research Capital.

What the product is

Epsilab builds what the AI field calls an environment. It is a simulated version of a job. An agent tries the job in the simulator. A grader marks what it did. The agent updates its weights. This is reinforcement learning. The quality of the environment decides the quality of the agent that comes out.

The company has its own pitch for why the models fail. Current frontier models "have a long way to go when deployed in the real world," in Epsilab's own words. Existing benchmarks are part of the problem. The launch page names TerminalBench 4.0 as an example. The page calls that benchmark broken. It cites "impossible criteria and unreliable graders." Epsilab says it grounds its own training gyms in real-world data. It says the gyms also draw on the judgement of top industry people. Those are the company's own claims. No outside review is named on the page.

Who the customer is

The company sells to frontier AI labs, not to end users. The launch page tags Epsilab as B2B. It lists machine learning, reinforcement learning and data labeling as its industry tags. A lab that is about to train the next version of an enterprise agent is the buyer. Epsilab sits one layer below the lab and one layer above the enterprise. Its pitch is that its environments mirror real office tasks. An agent trained on them should not break the moment a user tries a real one.

The company names no customer on the launch page. The amount of its YC funding is standard for a Fall 2026 batch and no other round is disclosed. The page gives no revenue number.

Why a buyer might listen

A software team buying an AI agent off the shelf has no say in how it was trained. A team paying a lab to build one does. Epsilab is a bet that the training environment is the right place for that team to push. The argument on the launch page is that a free benchmark is often wrong. A paid environment grounded in the work of named industry people is the correction.

The limit is the same one every new vendor in this space hits. Nobody outside Epsilab has yet tested the environments. A team considering a lab that trains on Epsilab's work can ask for the environment description and a sample task. A team funding a lab can ask whether it compared Epsilab's work to the free alternatives.

Source

Epsilab launch on Y Combinator and the product site at epsilab.com.