The infrastructure for self-improving agents
Documentation | Website | X | LinkedIn
BenchGen is a Synthetic Data Factory that creates a digital twin of your business, then uses that twin to simulate realistic environments, train RL agents, and generate unlimited synthetic trajectories. The result: agents that improve continuously on their own, grounded in your real business context.
- 🤖 Build with Agents — set up your agent and connect it to a simulation environment
- 📊 Evaluate with Eval — benchmark any model against curated environments, track performance, export failing cases
- 🏋️ Improve with Train — fine-tune on generated trajectories with managed LoRA runs, no GPU provisioning
- 🔄 Understand the Agent Learning Loop — how Simulate → Act → Capture → Measure → Generate → Improve compounds over time
- 📖 Read the Docs — full platform documentation
Every run through BenchGen produces better agents, richer trajectory data, and tighter evaluations — automatically.
- Simulate — your agent runs inside a digital twin of your enterprise: real APIs, databases, workflows, and policies
- Act — the agent uses real tools and integrations, the same ones it will use in production
- Capture — every decision, tool call, and state change is recorded in full as a high-fidelity trajectory
- Measure — outcomes are scored automatically using verifiable reward functions, no human review needed
- Generate Data — every run becomes a labeled training dataset; failures become improvement signal
- Improve — trajectories feed directly into LoRA fine-tuning; the improved model re-enters the loop
The loop compounds. More environments, more runs, better data, smarter agents.
We welcome contributions of all kinds — environments, datasets, examples, and feedback.
Start with any repo's README and check the open issues.
Questions? Find us on X or LinkedIn.
BenchGen builds open infrastructure to help teams ship better agents.
From evaluation environments to RL training and trajectory generation, we support developers across the entire agent lifecycle.