1 Startup · $40M · AI Agent Training Environments

What they're building

Bespoke Labs builds the reinforcement learning environments and surrounding infrastructure that frontier labs and enterprises use to train, evaluate, and improve long-horizon AI agents. Today's agents handle short tasks but fall apart over hours or days of autonomous work; to fix that, they need realistic environments to practice in and be measured against. Bespoke builds those environments with research techniques developed alongside the academic and open-source community, then helps customers measure and improve agents using methods like GEPA (Genetic-Pareto Agent Optimizer), which automates prompt and policy search. The team's open-research footprint is unusual for its size: core contributor to Terminal-Bench, one of the most cited agentic benchmarks, and creator of OpenThoughts, an open reasoning dataset downloaded more than 500,000 times and used by groups including Thinking Machines Lab, Meta, and Amazon.

Why this matters

RL environments are becoming the picks-and-shovels layer of the agent era: every lab and AI-native enterprise pushing long-horizon capability needs them, and most competitors staff the problem with contractors building shallow app-level simulations. Bespoke's bet is that environments are a research problem, not a data-labeling problem — and the syndicate suggests sophisticated buyers agree: Wing led the A, 8VC led the seed, and the angel list includes Jeff Dean plus operators from Anthropic, OpenAI, and Meta. Wing's Peter Wagner framed the thesis as a new generation of data and training infrastructure being required as labs push long-horizon agents. Founded in 2024 by CEO Mahesh Sathiamoorthy and Chief Science Officer Alex Dimakis (a UC Berkeley professor), the company already has real distribution through its open-source work, which functions as both credibility and inbound pipeline.