We are excited to partner with Mahesh Sathiamoorthy, Alex Dimakis, and the entire Bespoke Labs team, who today announced raising $40M across their seed and Series A rounds.
Every wave has a moment when the foundational infrastructure becomes obvious in hindsight.
The biggest breakthroughs in AI have almost always been unlocked by a new data asset or benchmark, not a flashier model. That was true for deep learning. It was true for LLMs. It will be true for agents.
Environments will define that next wave. Here is why.
Agents are unreliable. That single fact is the biggest thing standing between the demos people see today and the coworker-grade agents the industry promises. Compute is plentiful. Reinforcement learning infrastructure is commoditizing. Base models are improving on a known cadence. What will not be democratized, and what ultimately determines whether an agent can be trusted in production, is the environment it learns in.
Mahesh and Alex saw this before most, and the proof is already in the work. OpenThoughts has become one of the most widely used open reasoning datasets in the field. Terminal-Bench is how the world’s leading frontier labs benchmark agentic coding, and Bespoke Labs is a core contributor. GEPA is being deployed inside enterprise environments as a practical, data-driven alternative to prompt engineering. This is a team that ships things people use, in the open, with the research community.
We’ve entered the agentic era. The next frontier is the data, environments, and evaluation systems that determine how capable and reliable AI agents become in production. Bespoke Labs is building that infrastructure, and we couldn’t be more excited to be in their corner.
Models start the wave. Infrastructure defines the winners.
