Bespoke Labs Raises $40M to Build Training Grounds for Reliable AI Agents
Bespoke Labs has raised $40 million from Wing VC, 8VC, and angels working at Anthropic, OpenAI, and Meta to develop simulated environments where AI agents learn complex, real-world tasks. The company believes that improved training grounds, not larger models, will determine which AI agents are successful in production.
July 7, 2026 – 1:31 pm
AI agents can perform tasks like writing code and answering questions, but they still struggle with long, messy jobs. Bespoke Labs, a Mountain View startup, has secured $40 million to create training grounds that address this issue.
The funding round was led by Wing VC with support from 8VC. The list of investors includes angels affiliated with Anthropic, OpenAI, and Meta, as well as Google DeepMind’s Jeff Dean and dbt Labs’ Tristan Handy.
Practice Grounds for AI Agents
Today’s AI agents are capable but unreliable, performing well on short tasks but struggling with longer, more complex jobs that span hours or days. Bespoke Labs takes a different approach: they believe the solution lies not in bigger models but in better training environments.
The company creates simulated versions of real-world workplaces, complete with large codebases, microservices, logs, support tickets, email, and Slack threads. Agents train within these environments, learning multi-step workflows that reflect actual work scenarios. Bespoke then assists customers in optimizing these agents using an in-house optimizer called GEPA, which enables faster improvement than manual tuning.
A Research-Focused Approach
Founded in 2024 by CEO Mahesh Sathiamoorthy and chief scientist Alex Dimakis, Bespoke Labs has a lean academic team that contributes to projects like Terminal-Bench, a widely cited test of agent skill, and OpenThoughts, an open reasoning dataset downloaded over 500,000 times by labs including Meta and Amazon.
Instead of outsourcing environment creation, Bespoke positions itself as a research lab that develops and sells the training infrastructure.
The Market Opportunity
Bespoke Labs has timed its launch strategically. Independent tests indicate that the length of tasks AI agents can reliably complete doubles approximately every seven months (as noted by METR) or even faster, according to some analyses. This rapid progress means that environments must also evolve at a similar pace, which is precisely what Bespoke offers.
The company faces competition from various angles: self-learning agents, stress-testing, testing, and benchmarking firms, as well as companies focused on the economics of running AI agents at scale. Bespoke argues that the training ground, not the model, will be the deciding factor in which AI agents reach production.
Whether better training environments ultimately outperform larger models remains to be seen, but the answer will significantly impact the survival of these companies during the next funding cycle.