Bespoke Labs raises $40M to train agents
Bespoke Labs raised $40M (Wing VC) to build RL environments for training reliable AI agents, with angels from Anthropic, OpenAI, and Meta.
Bespoke Labs announced $40 million in total funding on July 6, split between an $8.25 million seed led by 8VC and a $31.75 million Series A led by Wing VC. Mayfield and The House Fund also participated in the Series A. Angel investors include people from Anthropic, OpenAI, and Meta.
The company was co-founded by Mahesh Sathiamoorthy, CEO, who productionized TPU usage for YouTube’s recommender systems at Google and introduced Semantic IDs and Generative Retrieval to recommender research, and Alex Dimakis, Chief Science Officer and former UT Austin faculty.
What Bespoke Labs builds
Bespoke Labs builds reinforcement learning environments and post-training infrastructure for AI agents. The environments simulate realistic enterprise settings: codebases, microservices, logs, tickets, email, and Slack threads. The premise is that agents trained in toy environments behave well in toy environments and fail when deployed against a real production system.
Their published products include OpenThoughts, an open reasoning dataset with over 500,000 downloads used by teams at Meta, Amazon, and AI2; Terminal-Bench, an agentic coding benchmark confirmed used by Anthropic, OpenAI, and Google DeepMind; and GEPA, a genetic-Pareto evolutionary optimizer for automated prompt and policy optimization.
The central thesis
Sathiamoorthy has stated the thesis directly: “The environment that an agent learns in is the only important component that is not going to be democratized and thus will ultimately determine whether the agent is reliable enough to be trusted in production.”
This is a bet that model weights are becoming a commodity and that the training environment is what will differentiate reliable agents from unreliable ones at scale. It is a similar logic to why synthetic data quality became a competitive moat before model quality itself.
What this means for operators
Most operators who run AI agents in production do not train them; they prompt them and hook them to APIs. The Bespoke Labs thesis does not change that in the near term. What it does change is the question to ask when evaluating which agent framework or foundation model to deploy: where and how were the agents in this framework trained, and does the training environment resemble your production environment?
The JADEPUFFER agentic ransomware brief from yesterday illustrated the practical failure mode: an agent that adapted its behavior in a live production environment it was not trained for. The verification habit guide covers the corresponding review layer for agents already in production.