Bespoke Labs Raises $40M to Build Environments that Enable Reliable Agents

We’re excited to announce our Seed and Series A financing rounds to build the world’s best frontier data research lab and support for the post-training needs of enterprises and frontier labs.

For the past two years, we've been heads-down at Bespoke doing world-class data curation research and shipping best-in-class reinforcement learning environments for training and optimizing AI agents. This funding lets us go a lot deeper on both.

Why we built Bespoke Labs

Agents are unreliable. That single fact limits how long they can operate autonomously, and it is the single biggest thing standing between the demos people see today and the coworker-grade agents the industry promises.

The best way to improve reliability right now is through post-training, and the best way to do post-training is with high-quality, complex, realistic environments. Compute is plentiful, reinforcement learning training infrastructure is rapidly commoditizing, and base models are improving on a known cadence. 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.

That's the bet. We built BespokeLabs to accelerate post-training, and we believe the way to do that is through the lens of data and environments. Our goal is to build the best data research lab in the world.

What we've shipped so far

• OpenThoughts is one of the most impactful open data curation projects of the past year. The open reasoning dataset has been downloaded hundreds of thousands of times, has powered work at Meta, Amazon, and AI2, and has been cited by researchers at Thinking Machines, Microsoft, Nvidia, and elsewhere. More than 100 researchers have used it to train models and push on LLM reasoning and reinforcement learning.

• Terminal-Bench is a leading agentic coding benchmark, and we're core contributors. Anthropic, OpenAI, and Google DeepMind use this benchmark to showcase the agentic abilities of frontier models.

• GEPA is a genetic‑search‑based agent optimizer that automates prompt and policy tuning using evolutionary algorithms. This data-driven method for optimizing agents has quickly become a state-of-the-art alternative to hand-tuned prompt engineering, and is already being used inside enterprise deployments.

These projects showcase that we work in the open, with the research community, on problems that matter.

Why we’re building a data research lab

Data matters a lot more than people give it credit for. Most of the fundamental breakthroughs in modern AI have been unlocked by a new dataset or benchmark. CIFAR helped create convolutional neural networks. ImageNet ushered in the era of deep learning. Snapshots of the open web made LLMs possible. OpenThoughts, in its own smaller way, has done the same for the current wave of reasoning research.

As reinforcement learning and agents collide, data research is taking on a new shape. The dataset is becoming a living, executable world that an agent acts inside of and gets graded on. Curating these reinforcement learning environments is now one of the most important, unsolved problems in AI.

That's the lab we're building.

The hard problems we are solving

This is the part of the post we care most about, because it's also our pitch to researchers and engineers thinking about where to spend the next few years of their careers.

Here are a few questions that keep us up at night:

• How do you measure the quality of an environment? To build reliable agents, “quality” has to be a science, with rigorous methods for predicting whether an environment will actually move a frontier model's capabilities before you deploy the compute to find out.

• How do you build worlds that are complex enough to support long-horizon agents? METR's data shows the length of tasks AI agents can complete is roughly doubling every seven months. Extrapolating that out, by the end of the decade we're talking about agents that need to remain coherent across multiday workflows. The environments that train those agents don't exist yet, so we have to invent them.

• How do you curate environments the way we curated reasoning data? Hiring experts to hand-build apps is not going to get the field to multiday autonomy. We need AI-driven pipelines for generating realistic worlds while knowing when to use humans, when to use models, and how to verify the results.

• How do you build living environments that look and feel like entire companies? Think Westworld (with more computers and fewer guns). The frontier of our work is figuring out how to simulate an agent’s true operating conditions (using Slack messages, emails, Jira tickets, microservices, logs, and other agents) so we can teach an agent to operate autonomously for a long time and accomplish economically valuable tasks.

We are doing deep research to answer these questions and building the reinforcement learning environments and infrastructure to enable reliable agents.

The infrastructure we are building

Underneath this research is a new platform for creating reinforcement learning environments, which includes:

• An environment engine that lets our expert network and our own systems compose realistic, multitool, multistep worlds far faster than hand-coding each one.

• A sandboxing and execution layer built for high throughput, low latency, and the kind of scale that post-training requires.

• An agent optimization layer that uses methods like GEPA and reinforcement learning on top of those environments to turn a customer's agent into one that works.

We work with frontier labs, neolabs, and enterprises to build this infrastructure. The work varies for each customer, but the result is always the same: better environments produce more reliable agents.

If you’re excited about frontier work, including building and researching high quality environments and the infrastructure they need, reach out here.

Our 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, and Meta, and others. Our $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.

Thank you to our team, our customers, our investors, and the broader research community that has built so much of this with us in the open!