Bespoke Labs
Series A
Bespoke Labs builds reinforcement-learning environments and infrastructure for production AI agents.
Founded
2024Bespoke Labs builds reinforcement-learning environments and infrastructure for production AI agents.
As AI agents are asked to perform longer, more autonomous tasks, the need for realistic training and evaluation environments is growing faster than the environments themselves can be built.
Frontier labs and AI-native enterprises are therefore looking for environment and data infrastructure that can keep pace with rapid model capability gains. Bespoke Labs argues that high-fidelity environments are the decisive lever for making agentic systems reliable enough for production use.
Bespoke Labs builds research-grade reinforcement-learning environments from real engineering artifacts, including production codebases, microservices, logs, support tickets, email threads, and Slack channels, rather than simpler contractor-built app-level simulations.
Its team combines research scientists and systems engineers, and it has established open-source credibility through OpenThoughts, which has been downloaded more than half a million times and used by Meta, Amazon, and Thinking Machines Lab, plus Terminal-Bench, cited by Anthropic, OpenAI, and Google DeepMind. The GEPA optimizer further differentiates the platform by automating prompt and policy search.
Bespoke Labs is a young company competing in a crowded market for agent evaluation and training infrastructure, where several well-funded rivals already offer competing platforms, datasets, and benchmarks.
Its model depends on frontier labs and large enterprises continuing to rely on externally built environments rather than building their own, and on its ability to synthesize environments that faithfully mirror the complexity of real production systems, which remains a difficult research and engineering problem.