01RL environments
RL environments
& infrastructure
Company-scale simulations with real codebases and microservices, allowing agents to master the complex, long-horizon workflows required for production.
Data Research LabSF Bay Area
We accelerate post-training and agent optimization
Bespoke Labs builds RL environments and infrastructure to let frontier labs and enterprises train long-horizon agents for production.
Trusted by Research Teams
- 200+teams using GEPA
in production - Trendingdataset on Hugging Face: OpenThoughts
- Cited"in Claude 4.5 release notes (Terminal-Bench)
What We Do
As AI agents move from demos to production, the bottleneck for reliability is no longer the model, it's the environment.
We are an applied AI research lab building the environment infrastructure for the agent revolution.
02Agent evaluation
Agent evaluation
& optimization
Automated prompt and policy searches via the GEPA optimizer achieve superior accuracy significantly faster than traditional, manual prompt engineering.
03Production‑grade RL
Production‑grade RL
& benchmarking
Collaborative research in RL, curation, and benchmarking ensures that training environments keep pace with the rapidly advancing frontier of model capabilities.
Featured Benchmark
AutoResearchExam
How far can an AI agent take a research idea in 24 hours?
Our benchmark tests agents on open-ended ML research tasks, measuring how quickly they improve, whether they sustain progress, and whether their solutions generalize to hidden test data.
- 29
- ML research tasks
- 24h
- Research per run
| Rank | Model | Score |
|---|---|---|
| 1 | Opus 5.5 | 0.609 |
| 2 | Claude Fable 5.1 | 0.602 |
| 3 | GPT-6 Astra | 0.600 |
| 4 | Argon | 0.594 |
| 5 | Claude Opus 5 | 0.579 |
| 6 | Grok 4.7 | 0.523 |
| 7 | GPT-5.6 Sol | 0.513 |
| 8 | Grok 4.6 | 0.493 |
| 9 | Gemini 3.8 Flash | 0.484 |
| 10 | Kimi K3 | 0.435 |
| 11 | Qwen3.8 Max | 0.423 |
| 12 | Muse Spark 1.3 | 0.379 |
AUARC measures hidden-test performance over time. Snapshot: . View live results.
Our Research Track Record
We ship products.
And we advance the field.
This combination of open research, benchmarks, and deployed optimization infrastructure makes Bespoke the environment partner of choice for the agentic era.
-
01
Openthoughts.ai
OpenThoughts
One of the best open reasoning datasets.
-
02
Read the paper
Terminal-Bench
First environment-based benchmark for agentic systems.
-
03
Read the paper
GEPA
Evolutionary algorithms for prompt optimization.
Get in Touch
Join us in building the foundation for the agentic world.
For Talent
Join the Team
Full-time researchers and engineers who want to do the most important work in AI, not just the most visible.
For Experts
Build with us
Domain experts interested in building RL environments and data pipelines.
For Labs & Enterprises
Work with us
Frontier labs or enterprises looking to improve agent reliability.