Bespoke Labs raises $40M Series A to build realistic training grounds for long-horizon AI agents
Bespoke Labs, a Mountain View AI infrastructure startup, has raised $40 million across seed and Series A rounds to scale its agent training platform. The Series A was led by Wing VC with participation from Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angels from Anthropic, OpenAI, and Meta. The seed round was led by 8VC with participation from Google DeepMind researcher Jeff Dean and other prominent engineers.
Bespoke Labs tackles a critical problem in agentic AI: agents can complete short tasks but fail over hours or days of autonomous operation. The company builds hyper-realistic synthetic environments—complete with codebases, microservices, Slack logs, and email—where agents learn long-horizon workflows using reinforcement learning. The team is research-focused, not contractor-driven, and contributed to Terminal-Bench, a leading agent capability benchmark cited by Anthropic, OpenAI, and Google DeepMind.
Capital will fund team expansion and infrastructure scaling. Benchmarks from METR show agent task complexity is doubling every seven months. Bespoke Labs is betting that better environments—not just larger models—will determine which agents reach production. For architects building agentic systems, this signals the market is consolidating around simulation and evaluation infrastructure as a distinct layer from the models themselves.