
BespokeLabsAi
Bespoke Labs builds reinforcement learning (RL) environments and infrastructure that let frontier labs and enterprises train AI agents for long, multi-step production workflows. Its Nimble models turn text into decisions, returning a probability for every allowed answer so code can act on confident calls and pass close ones to a person. More than 200 teams use its GEPA optimizer in production.
- AI Agents
- Artificial Intelligence
- Software Only
Funding
Raised to date
$7.3MRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.


Founders
Product
Problem
AI agents often work well in demos but fail in production, where reliability depends less on the model than on the environment it is trained and tested in. Teams lack realistic settings in which agents can practice long, multi-step workflows before they are trusted with real work.
Solution
Bespoke Labs runs an applied AI research lab that builds RL environments and infrastructure for training agents. It creates company-scale simulations with real codebases and microservices, so agents can practice the complex, long-horizon workflows they will meet in production. Its GEPA optimizer runs automated searches over prompts and policies, which the company says reach higher accuracy significantly faster than manual prompt engineering. The lab also publishes benchmarks, including AutoResearchExam, which measures how quickly an agent improves a hidden test score over 24-hour research runs. Separately, its Nimble models take a state and a set of allowed answers in one request and return a probability for each option.
Target Audience
Frontier AI labs and enterprises training AI agents, plus the AI engineers and researchers who build and evaluate them.
Features
- Train agents in simulated company environments built from real codebases and microservices.
- Run automated prompt and policy searches with the GEPA optimizer instead of tuning prompts by hand.
- Send one request to Nimble and get a probability for every allowed answer.
- Choose between a 9B model for high-volume routing and a 27B model for higher-stakes calls.
- Ask several questions in a single request, with true/false, multiple-choice and graded answer types.
- Test agents on AutoResearchExam, a 29-task benchmark of open-ended machine learning research.