Bespoke Labs raises $40M to build training environments that make AI agents reliable

Bespoke Labs raises $40M to train reliable AI agents

Bespoke Labs has raised $40 million from Wing VC, 8VC, and angels working at Anthropic, OpenAI, and Meta to build the simulated environments where AI agents learn long, messy, real-world tasks. Its bet: better training grounds, not bigger models, will decide which agents make it to production. AI agents can write code and answer questions, but they still fall apart on long, messy jobs. A Mountain View startup just raised $40M to build the training grounds that fix that. Bespoke Labs, which builds the environments that train and test AI agents, has raised $40 million, the company announced. The total spans a Series A led by Wing VC and an earlier seed led by 8VC. The backer list is unusually pointed: angels who work at Anthropic, OpenAI, and Meta, plus Google DeepMind's Jeff Dean and dbt Labs chief Tristan Handy. Practice grounds for agents Today's agents are capable but unreliable. They handle short tasks well. They still struggle to work on their own over hours or days, the way a colleague would. Bespoke's bet is that the fix is not a bigger model but a better place to practise. So it builds simulated versions of real firms: large codebases, microservices, logs, support tickets, email, and Slack threads. Agents train inside these worlds and learn the long, multi-step workflows that actually earn their keep. Bespoke then helps customers measure and tune them, using an in-house optimiser it calls GEPA to find better prompts and policies faster than hand-tuning allows. A research lab, not a contractor shop Founded in 2024 by CEO Mahesh Sathiamoorthy and chief scientist Alex Dimakis, the roughly 40-person team leans academic. It is a core contributor to Terminal-Bench, a widely cited test of agent skill. It also built OpenThoughts, an open reasoning dataset that labs including Meta and Amazon have downloaded more than 500,000 times. Rather than farm the work out to contractors, Bespoke treats environment-building as research and sells the infrastructure that results. Why it matters Bespoke timed this deliberately. Independent tests from METR find the length of tasks agents can reliably finish now doubles roughly every seven months, and Tech Funding News notes some analyses now put that closer to every four. Sustaining that curve means environments that grow harder just as fast, and that is exactly what Bespoke sells. Rivals crowd the field. They attack agent reliability from every angle, from self-learning agents to firms that stress-test, test, and benchmark them before they ship. Others chase the economics of running agents at scale. Bespoke is wagering that the training ground, not the model, decides which agents reach production. Whether better environments really beat bigger models is still an open question. The answer will help decide which of these companies survives the next funding cycle.

AI post-training startup Bespoke Labs raises $40M in funding

Bespoke Labs Inc., a startup working to streamline the post-training phase of artificial intelligence projects, has raised $40 million in funding. The company stated today that the capital arrived in two tranches. Bespoke Labs raised the bulk of the funds, $31.75 million, through a Series A round led by Wing VC. The firm was joined by Mayfield, The House Fund and employees at major tech firms such as Anthropic PBC. Bespoke Labs earlier raised $8.25 million from a consortium that included Google DeepMind chief scientist Jeff Dean. The workflow through which developers build a custom AI model comprises two main steps. The first is the pre-training phase, which equips the neural network with the core skills and knowledge it requires to answer prompts. The second phase, post-training, hones the model's reasoning skills. It can also provide improvements in other areas such as long-horizon task completion. Developers often carry out post-training using a method called reinforcement learning. The basic idea is to provide an AI with sample tasks similar to the work it will carry out in production. When the model completes a sample task correctly, it receives a "reward." The reward is a piece of data that adjusts the algorithm's configuration to boost its output quality. Reinforcement learning is carried out in virtual environments tailored to each project. For example, a productivity agent might be trained in a sandbox that simulates employee workstations. A coding agent, meanwhile, may require a simulated GitHub repository. Bespoke Labs offers a platform that makes it easier to create reinforcement learning environments. According to the company, the software generates simulations using automation workflows and input from a network of human experts. It claims that the platform does so significantly faster than traditional manual approaches. The platform runs the AI environments that it generates using what Bespoke Labs describes as a sandboxing layer. According to the company, the latter component helps minimize latency and boost throughput. The platform's third core component automatically optimizes the output quality of the AI models being trained. One of the technologies that it uses for the task is GEPA, an open-source project Bespoke Labs released last year. The software automates prompt engineering, the process of finding the specific requests and prompt formats that maximize an AI model's output quality. Reinforcement learning is not the sole focus of Bespoke Labs' open-source work. The company is also prioritizing another popular post-training method called supervised fine-tuning, or SFP. It works by providing AI models with a set of sample prompts and answers that they can use to refine their output. Assembling SFP question sets can be a highly time-consuming process. Last January, Bespoke Labs released a dataset called OpenThoughts that contains more than a million sample prompts and responses. The company says that OpenThoughts provides better post-training results than earlier SFT datasets. Bespoke Labs will use its newly raised capital to enhance its reinforcement learning platform and finance more AI data research. Photo of Bespoke Labs' founders: Bespoke Labs