The future of artificial intelligence, if you believe the current hype cycle, belongs to agents—software that can think through multi-step problems without constant human supervision. But there's a less glamorous challenge lurking beneath the surface: how do you actually train these systems to work reliably in the real world?
Bespoke Labs, a Mountain View startup barely two years old, thinks it has an answer. And on July 6, 2026, investors signaled they're willing to pay for it. The company announced $40 million in combined funding—a $31.3 million Series A led by Wing Venture Capital, following an $8.25 million seed round led by 8VC. (The total rounds to $40 million from $39.55 million.)
Founded in 2024 by Mahesh Sathiamoorthy and Alex Dimakis, Bespoke Labs builds what amounts to elaborate training grounds for AI: reinforcement learning environments where agents can practice handling complex, long-running tasks. Think less chatbot, more software that spends hours debugging a codebase or navigating enterprise workflows without breaking.
Environments as the New Bottleneck
Wing's investment thesis, laid out in a blog post accompanying the announcement, frames the problem in stark terms. The real constraint in advanced AI training isn't compute or even data anymore—it's environments. Specifically, the kind of containerized, automatically refreshing simulations that mirror production systems: real codebases, functioning microservices, verifiable success metrics.
"The industry is migrating from labeled datasets to environment-based training," Peter Wagner, a managing director at Wing, wrote. It's a shift that creates obvious demand for infrastructure, which is presumably why Wing wrote the check.
Bespoke Labs lists customers across developer tooling, financial software, and data platforms, plus contracts with what it calls "frontier labs"—the handful of companies racing to build cutting-edge AI models. Specific names are scarce, though the company recently posted about a technical collaboration with Credit Karma and Intuit. That appeared on LinkedIn just weeks before the funding announcement, perhaps not coincidentally.
The Pedigree Factor
Sathiamoorthy comes from nearly a decade at Google and Google DeepMind, where he led generative retrieval work for recommender systems. That research showed up at NeurIPS 2023. Wing emphasized his operational track record—more than 50 production systems launched, TPU infrastructure scaled—calling him a rare engineer who doesn't just publish papers but ships code that works at scale.
Dimakis brings a different kind of credibility. Now a professor at UC Berkeley's EECS department and an IEEE Fellow, he's been involved in early academic generative AI efforts and helped lead the DataComp benchmark initiative. His presence signals research ambitions, something visible in the company's ICLR 2026 paper trail and open-source contributions.
Together, they represent a blend the venture world loves right now: deep learning theory married to production engineering.
What They're Actually Building
The product suite breaks into four main pieces. Terminal-Bench, accepted to ICLR 2026, offers a benchmark for evaluating agents on developer workflows inside terminal environments. GEPA—a reflective optimizer for prompts and agent code—also landed at ICLR 2026 and, according to Wing's characterization, has been deployed by enterprises for production use.
Then there's the OpenThoughts dataset and OpenThinker models, open reasoning resources released earlier this year. Download figures here are slightly murky: Wing claims more than 100,000 downloads as of July, while Bespoke Labs' own website cites 10,000-plus. The discrepancy could reflect timing, measurement methods, or the usual fuzziness around open-source metrics. Either way, there's traction.
Rounding out the lineup is Bespoke-MiniCheck, a model designed to verify factual grounding in agent outputs—essentially, a fact-checker for AI systems prone to hallucination.
The Capital and the Crowd
Beyond Wing and 8VC, the investor roster reads like a who's who of AI-adjacent players. Mayfield, The House Fund, and dbt Labs CEO Tristan Handy participated, along with angels from Anthropic, OpenAI, and Meta. Jeff Dean—yes, that Jeff Dean—joined the seed round, as did Resolve AI CEO Spiros Xanthos and DevRev CEO Dheeraj Pandey.
Bespoke Labs says it will use the capital for research expansion, hiring, and business development. Standard language, but necessary: LinkedIn shows 62 employees, and job postings from earlier this year show open roles for reinforcement learning environment engineers and engagement managers.
The funding arrives at a moment when long-horizon agent capabilities are accelerating quickly. Research from METR, updated through early 2026, suggests the task length autonomous agents can reliably handle has doubled roughly every seven months since 2019. There's some indication that pace may have picked up since 2024, though causality is always tricky with these timelines.
That trajectory, real or projected, creates a plausible market for more sophisticated training infrastructure. Which is why a company founded just two years ago can attract backing from infrastructure VCs and veterans of frontier AI labs alike.
Valuation terms weren't disclosed. The company remains headquartered in Mountain View, with team members scattered across the Bay Area—par for the course in this era of hybrid work and distributed teams.
Whether Bespoke Labs can maintain its momentum as the agent landscape evolves remains an open question. But for now, at least, it's raised enough capital to find out.
