Mountain View, California-based Bespoke Labs announced on July 6, 2026 that it has raised $40 million to expand the reinforcement-learning environments and infrastructure it sells to frontier AI labs and enterprises training long-horizon autonomous agents. Wing Venture Capital led the $31.75 million Series A, joined by Mayfield, The House Fund, dbt Labs CEO Tristan Handy and individual angels working at Anthropic, OpenAI and Meta; the tranche is stacked on an earlier $8.25 million seed led by 8VC with Google DeepMind chief scientist Jeff Dean, Resolve AI CEO Spiros Xanthos and DevRev CEO Dheeraj Pandey among the backers.
Practice Worlds, Not Bigger Models
Founded in 2024 by CEO Mahesh Sathiamoorthy and Chief Science Officer Alex Dimakis, Bespoke Labs builds simulated enterprise environments — codebases, Slack threads, tickets, emails, microservices and operational logs — that behave like real companies so agents can practice long-horizon tasks before they are deployed. Independent METR benchmarks show the length of tasks AI agents can reliably complete is now doubling roughly every seven months, and analysts argue that pace is bottlenecked by training environments rather than model size. Bespoke's platform pairs an environment builder with a low-latency sandbox and the open-source GEPA optimizer, which automates prompt-and-policy search to accelerate post-training reliability gains far beyond hand-tuned prompt engineering.
Racing Scale AI and Poolside for the Agent Stack
The round comes as capital pours into every layer of the agent stack. Meta paid $14.3 billion to take a stake in labelled-data leader Scale AI, coding-model startup Poolside raised $500 million, and Reflection AI banked $130 million for autonomous software engineering agents. Bespoke is betting the practice grounds where agents train are the next durable moat, positioning itself as neutral infrastructure rather than another foundation-model house. Investor Peter Wagner, founding partner of Wing VC, said "a new generation of data and training infrastructure is required" as frontier labs push long-horizon capabilities. The startup's team already contributes to Terminal-Bench, one of the most cited agent benchmarks, and the OpenThoughts reasoning dataset, downloaded more than 500,000 times and used by Thinking Machines Lab, Meta and Amazon.
What the Capital Buys
Bespoke says the funding will go toward hiring research scientists and engineers, expanding its environment-generation pipelines and snapshotting real-world infrastructure into training environments. The company positions itself as a research lab first, competing with cheaper contractor-built environments on depth and difficulty rather than volume. The bet lines up with the broader shift analysts flagged this week as Chamath's 8090 raised $135M for an AI software factory and after Anthropic pushed Claude into agentic Google Workspace flows. The deal also arrived the same week Anthropic inked a $19 billion, 20-year lease with TeraWulf for AI data-center capacity, underscoring how much the market values every rung of the agent training ladder.
Reporting based on coverage from Business Wire, SiliconANGLE and Tech Funding News.