Bespoke Labs announced on July 6th that it has raised $40 million across seed and Series A financing to build simulated training environments where AI agents can practice enterprise work before they are allowed into live systems.
The Mountain View AI infrastructure lab, co-founded by Mahesh Sathiamoorthy and Alex Dimakis, disclosed the financing in a Business Wire announcement after the news was flagged by Aligned News. Bespoke Labs said the Series A was led by Wing VC, with participation from Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angels from Anthropic, OpenAI and Meta. The seed round was led by 8VC, with participation from Jeff Dean, Resolve AI CEO Spiros Xanthos, DevRev CEO Dheeraj Pandey and others.
Bespoke Labs did not disclose the split between the seed and Series A, the valuation, revenue, paying customer count or named customers. That leaves the $40 million total as the only hard financing number. The capital, Bespoke Labs said, will go toward research hiring, environment-building infrastructure and business development.
The bet is direct: if agents are going to do multi-step work inside a company, they need a place to fail where the failure does not touch production. Bespoke Labs says its environments mimic large codebases, microservices, logs, tickets, email, Slack and other internal workflow context, so agents can learn and be evaluated on long-horizon tasks with repeatable starting states and measurable outcomes.
Sathiamoorthy is a relevant founder for that problem. His own site says he started Bespoke Labs in 2024 after working as a staff software engineer at Google DeepMind and Google Brain, where he worked on productionizing TPUs for YouTube recommender systems and on semantic IDs and generative retrieval for recommenders. His path runs from IIT Kharagpur to USC to Google, then into Bespoke Labs, where he says the focus is RL environment curation and data curation.
Dimakis gives Bespoke Labs a second technical center of gravity. UC Berkeley lists him as a professor in Electrical Engineering and Computer Sciences, with research interests that include generative AI, information theory and machine learning. Berkeley's faculty page says he has published over 150 papers, received a PhD from UC Berkeley in 2008 and is an IEEE Fellow. Bespoke Labs lists him as co-founder and chief science officer.
The product is the practice field
Bespoke Labs' pitch rests on a shift already visible across agent infrastructure: training data is no longer only documents, labels and preference rankings. The work now includes interactive systems that can reset, record trajectories, provide rewards and verify whether a task actually changed the underlying state.
Bespoke Labs says it builds company-scale RL environments and infrastructure for frontier labs and enterprises. On its own site, Bespoke Labs describes three core pieces: RL environments and infrastructure, agent evaluation and optimization, and production-grade RL and benchmarking. Bespoke Labs also claims more than 200 teams use GEPA, its prompt and policy optimizer, in production, and says OpenThoughts receives over 10,000 monthly downloads on Hugging Face.
Those are Bespoke Labs' numbers. The more verifiable public footprint is its open-source work. Curator, Bespoke Labs' synthetic data curation project for post-training and structured data extraction, had about 1,700 GitHub stars when checked and lists a January 2025 launch in its repository history. The Bespoke Labs docs describe Curator as an open-source project for generating high-quality synthetic data at scale for fine-tuning and structured extraction.
Bespoke Labs is also positioning its research artifacts as distribution. The July 6th announcement says Bespoke Labs is a core contributor to Terminal-Bench and is behind OpenThoughts, an open reasoning dataset that Bespoke Labs says has been downloaded over 500,000 times and used by groups including Thinking Machines Lab, Meta and Amazon. That claim comes from Bespoke Labs' release, and the public customer evidence remains thin: Bespoke Labs has not publicly named enterprise customers, published pricing, disclosed security certifications or released third-party production case studies.
A crowded category by July
Bespoke Labs is not alone in seeing environments as the next agent bottleneck. Scale AI markets RL Environments for long-horizon professional workflows, including simulated APIs, MCP servers, GUIs, expert-curated artifacts and automated verifiers. Scale introduced the product in a Feb. 27th blog post, saying nearly half of its new data training projects involved reinforcement learning environments.
G2i is selling RL environments for complex apps across UI, data and APIs, with claims around repeated runs, resets, checkpoints, telemetry and batch execution. Prime Intellect launched Lab on Feb. 10th as a platform that combines environments, hosted training and hosted evaluations for agentic post-training. Daytona, which raised a $24 million Series A on Feb. 5th, is adjacent infrastructure: programmable sandbox computers where agents can execute code, branch, snapshot and run safely.
The competitive pressure matters because Bespoke Labs is trying to own the hardest part of the agent stack without yet showing public customer depth. A basic sandbox can isolate code execution. A useful RL environment has to recreate enough of a real business process to teach the agent something transferable, while producing rewards and verifiers that cannot be gamed by plausible text. That requires research, data curation, domain expertise and software infrastructure in the same package.
In the release, the team emphasized high-quality environments as key to optimizing and developing agents, and described work with labs on reinforcement learning, environment curation, benchmarking and long-horizon modeling so environments keep pace with agent capabilities.
The financing gives Bespoke Labs time to prove that approach with customers whose names and benchmarks can be checked. For now, investors are underwriting the founding team's research credibility and the market's move from chat agents toward agents that use tools, operate across systems and need measurable training loops. The next proof point will be harder than a funding announcement: agents trained in Bespoke Labs environments will need to work better when they leave the simulation.