Bespoke Labs raises $40M to build training environments that make AI agents reliable

Bespoke Labs Secures $40M Funding to Transform AI Agent Training

Bespoke Labs has raised $40 million to build the infrastructure that trains AI agents to handle complex, multi-step tasks that unfold over hours or days 1

1

. The AI post-training startup announced that the capital arrived in two tranches: a $31.75 million Series A led by Wing VC, with participation from Mayfield, The House Fund, and employees at major AI labs, plus an earlier $8.25 million seed round led by 8VC that included Google DeepMind chief scientist Jeff Dean 2

2

. The backer list features angels working at Anthropic, OpenAI, and Meta, alongside dbt Labs chief Tristan Handy 1

1

.

Building Simulated Environments for Training AI Agents

Founded in 2024 by CEO Mahesh Sathiamoorthy and chief scientist Alex Dimakis, Bespoke Labs operates on a core thesis: better training grounds, not bigger models, will determine which AI agents make it to production 1

1

. The company builds simulated versions of real firms complete with large codebases, microservices, logs, support tickets, email, and Slack threads where agents can practice long, multi-step workflows 1

1

. The platform generates these simulations using automation workflows and input from a network of human experts, significantly faster than traditional manual approaches 2

2

.

Advanced Reinforcement Learning Platform with Post-Training Optimization

Bespoke Labs offers a reinforcement learning platform that streamlines the post-training phase of AI projects, the critical step that hones AI model reasoning and improves long-horizon task completion 2

2

. The platform runs AI environments using a sandboxing layer designed to minimize latency and boost throughput 2

2

. To optimize output quality, Bespoke Labs built GEPA, an in-house optimizer that finds better prompts and policies faster than hand-tuning allows 1

1

. GEPA automates prompt engineering, identifying the specific requests and formats that maximize an AI model's performance 2

2

.

Research-Driven Approach to Creating Reliable AI Agents

The roughly 40-person team treats environment-building as AI data research rather than contracting work 1

1

. Bespoke Labs serves as a core contributor to Terminal-Bench, a widely cited test of agent skill 1

1

. The company also released OpenThoughts, an open reasoning dataset downloaded more than 500,000 times by labs including Meta and Amazon 1

1

. OpenThoughts contains over a million sample prompts and responses designed for supervised fine-tuning, providing better post-training results than earlier datasets 2

2

.

Why This Matters for AI Development

Today's AI agents handle short tasks well but struggle to work autonomously over extended periods the way a colleague would 1

1

. Independent tests from METR find the length of tasks agents can reliably finish now doubles roughly every seven months, with some analyses putting that closer to every four months 1

1

. Sustaining this curve demands simulated environments for training AI agents that grow harder just as fast, positioning Bespoke Labs at a critical inflection point. The company will use its newly raised capital to enhance its reinforcement learning platform and finance more AI data research 2

2

. While rivals attack agent reliability from multiple angles including self-learning systems and stress-testing platforms, Bespoke Labs is wagering that the training ground itself determines which agents reach production 1

1

.