Bespoke Labs

Bespoke Labs builds reinforcement learning environments and infrastructure that help AI labs and enterprises train and evaluate agents for complex, multistep work. Its company-scale simulations incorporate real codebases and microservices, letting agents practice long workflows and enterprises test behavior in environments that mirror their systems. Its infrastructure combines environment creation, sandboxed execution, and automated prompt and policy optimization.

The company was co-founded by CEO Mahesh Sathiamoorthy, previously at Google and DeepMind, and Chief Scientist Alex Dimakis, a machine learning professor at UC Berkeley. Its collaborative research includes OpenThoughts, which supplies reasoning datasets and models alongside publicly available data generation, training, and evaluation code. Another contribution, GEPA, uses natural-language reflection on agent execution traces to diagnose problems, propose and test prompt changes, and combine improvements through evolutionary search. These projects address both the data used to train reasoning models and the instructions that guide agents.

Bespoke works with frontier labs on training infrastructure and with enterprises on optimizing production agents. As of August 2026, the company reported more than 200 teams using GEPA in production. In 2026, it announced $40 million raised across seed and Series A financing, with Wing VC leading the Series A.