How Bespoke Labs is Building Environments that Enable Reliable Agents

Bespoke Labs, creating AI tools for data curation and post-training LLMs, led by Mahesh Sathiamoorthy and Alex Dimakis, has announced its Seed and Series A. Series A round was led by Wing VC, with participation from Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angels from Anthropic, OpenAI, Meta, and others. The $8.25 million seed round was led by 8VC, with participation from Jeff Dean, Resolve AI CEO Spiros Xanthos, DevRev CEO Dheeraj Pandey, and others.

Why Bespoke Labs?

Reliable AI agents require more than better models, they require better training environments. While compute, training infrastructure, and foundation models continue to improve, the quality of the environments used for post-training is becoming the key factor in building agents that can operate reliably in production.

Bespoke Labs was founded on that belief, with a focus on accelerating post-training through high-quality data and realistic environments to build more dependable AI agents.

What They Have Built?

Bespoke Labs has contributed several widely adopted open-source AI projects, including OpenThoughts, an open reasoning dataset used by researchers and organizations such as Meta, Amazon, AI2, Microsoft, Nvidia, and Thinking Machines.

The team is also a core contributor to Terminal-Bench, a leading benchmark for agentic coding used by Anthropic, OpenAI, and Google DeepMind, and developed GEPA, an optimization framework that automates prompt and policy tuning for AI agents.

Together, these projects reflect Bespoke Labs' commitment to advancing open AI research and building tools that improve the reliability of AI agents.

Why a Data Research Lab?

Advances in AI have often been driven by breakthroughs in datasets and benchmarks. As reinforcement learning and AI agents evolve, static datasets are giving way to executable environments where agents learn, act, and are evaluated.

Bespoke Labs is building a data research lab focused on creating these next-generation training environments, with the goal of improving the reliability and performance of AI agents.

The Hard Problems.

Bespoke Labs is focused on some of the most challenging questions in AI: how to measure the quality of training environments, build realistic worlds that support long-horizon agents, and automate the creation of those environments at scale.

The company is also developing "living" environments that simulate real enterprise workflows, including emails, messaging platforms, tickets, logs, and software systems, allowing AI agents to learn in conditions that closely mirror production. The goal is to build the reinforcement learning infrastructure needed for reliable, long-running autonomous agents.

The Infrastructure.

Bespoke Labs is building a platform for creating reinforcement learning environments that includes an environment engine for generating realistic multi-step workflows, a high-performance sandbox and execution layer, and an optimization framework that improves AI agents using techniques such as reinforcement learning and GEPA.

The platform is designed for frontier AI labs, emerging model developers, and enterprises, with the goal of improving agent reliability through better training environments.