Epsilab builds training environments around the realities of enterprise work

Epsilab is building reinforcement-learning environments intended to reflect how enterprise work actually gets done. The Fall 2026 YC company describes itself as an applied AI research lab working on long-horizon tasks, where an agent must complete a workflow rather than produce a single convincing answer.

Its focus is the training environment around the model. Epsilab says it grounds these environments in real-world data and the judgment of experienced professionals, with the aim of improving agents' ability to handle work outside a benchmark. The company says it works directly with frontier labs, but does not name those partners in its launch.

The underlying concern is that a model learns against the tasks and feedback it is given. If an environment misrepresents the work, or its grading fails to distinguish success from failure, a rising score may say less about useful performance than it appears to. Epsilab's product thesis is that realistic tasks and reliable evaluation deserve as much attention as the training process itself.

Watch Epsilab's original launch video.

In its announcement, Epsilab argues that creating challenging environments increasingly requires domain expertise alongside engineering, operations and research. Rather than make tasks artificially difficult, it wants the challenge to come from the actual demands of professional work. It describes maintaining that level of fidelity as costly and central to its approach.

The company's public YC profile also lists real-world environment simulation, long-horizon workflows and multi-player agents among its research and evaluation interests. Those categories indicate a broader ambition than labeling isolated examples. They do not, however, establish which industries are already covered or which environments are commercially available.

Epsilab's founders have worked across AI products, research and finance. Joanne Wang previously built Copper Health and worked in investment banking and private equity, according to the launch. Fengjiao Peng previously co-founded Shortbread and worked in machine-learning research, including at MIT Media Lab. Evan Hu was Shortbread's co-founder and CTO and studied at UC Berkeley. Matthew Olsen previously worked on quantitative infrastructure at Tower Research Capital.

That combination is relevant to the company's emphasis on both technical execution and professional judgment. An environment needs software that can represent a workflow, but it also needs a credible account of what a good outcome looks like. Epsilab says it partners with high-performing professionals to bring that expertise into the training process.

The launch presents a direction and a team, rather than a quantified performance case. It does not disclose pricing, named enterprise customers, independently measured model improvements or a detailed catalog of training environments. Its claims about realism therefore remain a proposition to evaluate, not evidence that an agent trained in its environments is ready for unsupervised enterprise deployment.

For frontier labs, the potential value is a better connection between progress during training and performance on the work agents are expected to handle. That connection becomes harder to measure as tasks stretch across more steps and depend on judgment that cannot be captured by a single easy-to-check answer.