BenchGen

BenchGen is an AI agent evaluation and benchmarking platform that closes the gap between development and production. It provides infrastructure to simulate real operational environments, run agents through multi-step workflows, score agent quality across five measurable dimensions (tool-call accuracy, skill coverage, goal completion, memory utilisation, regression stability), and export clean trajectory data for LoRA fine-tuning on open-weight models.

Every BenchGen environment is built around RLVR (Reinforcement Learning with Verifiable Rewards): each task carries a machine-checkable reward signal, so evaluation results are objective and reproducible, not opinion-based. This makes the full loop possible - benchmark > evaluate > fine-tune > re-evaluate - without relying on human annotation at every step.

BenchGen auto-detects your agent directory, reads trajectory files from live operation, and produces a quality report in under five minutes. Teams use it to catch regressions after model or skill updates before users do, and to build the filtered training datasets needed for continuous fine-tuning.