Bespoke Labs

Bespoke Labs is an applied AI research lab focused on building the environments and infrastructure required to make AI agents reliable for production use. Founded in 2024 by CEO Mahesh Sathiamoorthy, a former research engineer at Google DeepMind, and Chief Science Officer Alex Dimakis, a professor at UC Berkeley, the company is based in Mountain View, California. The founding team combines experience from frontier labs, major tech companies, and leading academic institutions.

The company addresses the critical challenge of AI agent unreliability in long-horizon, complex tasks. While agents can perform short tasks, they often fail when operating autonomously over extended periods in real-world scenarios. Bespoke Labs' core offering is a platform that creates realistic, company-scale reinforcement learning (RL) environments. These simulated environments mimic actual company operations, incorporating large codebases, microservices, logs, support tickets, and communication platforms like email and Slack. This allows AI agents to train on, learn from, and be measured against the complex, multi-step workflows they will encounter in production.

Bespoke Labs serves frontier AI labs and enterprises, providing the infrastructure to train, evaluate, and improve long-horizon agents. The business model centers on providing this sophisticated training infrastructure as a service, positioning itself as a key partner in the agentic AI era. The platform features an environment engine for composing realistic worlds, a high-throughput sandboxing and execution layer, and an agent optimization layer that uses proprietary techniques like the Genetic-Pareto Agent Optimizer (GEPA). GEPA automates prompt and policy tuning through evolutionary algorithms, achieving higher accuracy faster than manual engineering.

The firm takes a research-first approach, actively contributing to open-source projects. It is a core contributor to Terminal-Bench, a widely used benchmark for agentic coding skills cited by labs like Anthropic, OpenAI, and Google DeepMind. The company also created OpenThoughts, an open reasoning dataset with over 500,000 downloads used by companies like Meta and Amazon. In July 2026, Bespoke Labs announced it had raised $40 million through a seed round led by 8VC and a Series A round led by Wing VC, with participation from notable investors including Mayfield, The House Fund, and angel investors from Anthropic, OpenAI, and Meta.

Keywords: AI agent training, reinforcement learning environments, agent reliability, long-horizon agents, data curation, agent evaluation, agentic AI, synthetic data generation, prompt optimization, GEPA optimizer, Terminal-Bench, OpenThoughts, AI infrastructure, machine learning, enterprise AI, AI development tools, model post-training, AI simulation, agentic workflows, company-scale simulation