Bespoke Labs’ $40M Bet on Building the Digital Gyms for AI Agents

MOUNTAIN VIEW, CA – July 06, 2026 – In a landscape saturated with headlines about ever-larger AI models, a quieter but arguably more critical part of the ecosystem just received a major vote of confidence. Bespoke Labs, a startup focused on creating training environments for AI agents, announced today it has secured $40 million in funding. The capital, raised across a Series A led by Wing VC and a seed round led by 8VC, signals a growing recognition of a fundamental bottleneck in artificial intelligence: reliability.

While today’s AI agents can write code, answer questions, and perform isolated tasks with impressive skill, their ability to operate autonomously over hours or days—emulating a human coworker—remains fraught with error. Bespoke Labs is not building the agents themselves; instead, it's building their virtual bootcamps. The new funding, which includes participation from Mayfield, The House Fund, and prominent angel investors from Anthropic, OpenAI, and Meta, will be used to expand its research team and scale the construction of these sophisticated digital proving grounds.

“As frontier labs and AI-native enterprises push the boundaries of long horizon agentic capabilities, a new generation of data and training infrastructure is required,” said Peter Wagner, Founding Partner of Wing Venture Capital. This statement cuts to the core of the issue. The race for AI supremacy is no longer just about model size, but about creating agents that can be trusted to execute complex, multi-step tasks in the messy reality of the corporate world.

The Reliability Gap: From Brittle Tools to Autonomous Partners

The central problem Bespoke Labs aims to solve is the “production readiness gap.” While demos are dazzling, real-world reliability for AI agents hovers around 80%, a figure that would be unacceptable for any critical business function. Errors in one step can cascade, outputs can be non-deterministic, and integrating agents into legacy systems is a monumental challenge. This is the difference between a clever parlor trick and a dependable digital employee.

To bridge this gap, agents need practice. According to independent benchmarks from METR, the length of tasks AI agents can reliably complete has been doubling roughly every seven months—an exponential trend that has held for six years. Sustaining this trajectory requires training environments that grow in complexity at the same pace. This is where Bespoke Labs plants its flag.

The company designs and builds what it calls “company-scale” reinforcement learning environments. These are not simple app-level simulations. They are high-fidelity digital replicas of real businesses, complete with large codebases, microservices, realistic logs, ticketing systems, email traffic, and even simulated Slack channels. Within these digital ecosystems, agents can learn the long-horizon workflows that are economically meaningful, moving beyond simple automation to tackle genuine business problems.

“Frontier labs, enterprises, and all organizations relying on reliable agents need access to high-quality environments,” explained Mahesh Sathiamoorthy, co-founder and CEO of Bespoke Labs. “This is the critical piece needed to optimize and develop agents.”

A Research-First Philosophy in a Hype-Driven Market

Bespoke Labs distinguishes itself with a “research-first” approach, a stark contrast to competitors who, according to the company, often use contractors to build simpler environments. The team is composed of research scientists and engineers dedicated to pushing the frontier of environment curation itself.

This academic rigor is evident in their contributions to the field. Bespoke Labs is a core contributor to Terminal-Bench, a widely cited open-source benchmark that evaluates an agent's ability to perform real-world tasks in a command-line environment, such as compiling code or configuring a server. They are also the team behind OpenThoughts, an open reasoning dataset downloaded over 500,000 times and used by research groups at Meta and Amazon to advance reasoning models.

Internally, the company develops novel techniques for measuring and improving agent performance. One such method is the Genetic-Pareto Agent Optimizer (GEPA), a system that automates the laborious process of prompt engineering. Instead of having humans manually tweak instructions, GEPA uses LLMs to analyze an agent's failures and systematically propose improvements, reaching higher accuracy much faster. This scientific, methodical approach to building trustworthy AI is what attracted investors like Wing VC, whose founding partner Peter Wagner noted that the founders “deeply understand the needs of leading AI researchers.”

Unlocking the AI Coworker and the $52 Billion Market

The ultimate vision powered by this work is the advent of the true AI coworker. As agents become more reliable, they can transition from being peripheral tools to core team members, capable of autonomously handling tasks in customer support, procurement, software development, and supply chain management. This shift could unlock immense productivity, freeing human talent to focus on more strategic, creative initiatives.

The market is poised for this transformation. Projections show the AI agent market growing from $7.8 billion in 2025 to over $52 billion by 2030. However, significant hurdles remain. Integrating these advanced agents into complex enterprise systems without causing disruption is a major challenge, alongside ensuring data privacy, security, and regulatory compliance.

Bespoke Labs' focus on creating robust, sandboxed environments is a direct answer to these challenges. By providing a safe and realistic space for agents to train, fail, and learn, the company is building the essential infrastructure needed for the next wave of AI adoption. Their work is a reminder that before an AI can run a company, it must first graduate from a school that looks exactly like one.