Bespoke Labs Raises $40M to Build the Gyms Where Agents Grow Up
Everyone's racing to build a bigger model. Bespoke Labs just raised $40 million arguing the bottleneck is somewhere else entirely: the training ground.
The Series A, announced July 6, was led by Wing VC, with Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angels from inside Anthropic, OpenAI and Meta all piling in. The pitch is that reliable agents don't come from scale, they come from environments, realistic and messy simulations of the actual workplace. Bespoke builds fake-but-faithful codebases, microservices, and communication logs where a long-horizon agent can be trained and stress-tested before it touches anything real. Think of it as the flight simulator the industry skipped straight past on its way to shipping autopilots.
The team has the receipts. Founders Mahesh Sathiamoorthy and Alex Dimakis are core contributors to Terminal-Bench, one of the most-cited agent benchmarks, and built OpenThoughts, a reasoning dataset downloaded over half a million times and used by Thinking Machines Lab, Meta and Amazon. These aren't tourists. They've been quietly supplying the eval and data plumbing everyone else trains on.
The framing that matters is a single question: can better environments beat bigger models? That's a direct shot at the scale-is-all-you-need crowd, and it's the same bet showing up all over this week's research, from multi-platform distillation to continual-learning environments to harness-over-fine-tuning. The whole field is quietly discovering that where you train the agent matters as much as how big it is. Bespoke wants to be the company that sells the where.
Details: bespokelabs.ai
← Back to all articles
The Series A, announced July 6, was led by Wing VC, with Mayfield, The House Fund, dbt Labs CEO Tristan Handy, and angels from inside Anthropic, OpenAI and Meta all piling in. The pitch is that reliable agents don't come from scale, they come from environments, realistic and messy simulations of the actual workplace. Bespoke builds fake-but-faithful codebases, microservices, and communication logs where a long-horizon agent can be trained and stress-tested before it touches anything real. Think of it as the flight simulator the industry skipped straight past on its way to shipping autopilots.
The team has the receipts. Founders Mahesh Sathiamoorthy and Alex Dimakis are core contributors to Terminal-Bench, one of the most-cited agent benchmarks, and built OpenThoughts, a reasoning dataset downloaded over half a million times and used by Thinking Machines Lab, Meta and Amazon. These aren't tourists. They've been quietly supplying the eval and data plumbing everyone else trains on.
The framing that matters is a single question: can better environments beat bigger models? That's a direct shot at the scale-is-all-you-need crowd, and it's the same bet showing up all over this week's research, from multi-platform distillation to continual-learning environments to harness-over-fine-tuning. The whole field is quietly discovering that where you train the agent matters as much as how big it is. Bespoke wants to be the company that sells the where.
Details: bespokelabs.ai
Comments