What it is
Synthesizes executable business worlds from high-level scenarios, then uses them to train agents and learn environment authoring itself. Instead of building environments around benchmark tasks, it builds persistent, verifiable business worlds from scenarios; tasks emerge later from the world, and construction traces supervise world authoring.
Key results
- Qwen3.5-4B: 12.3 -> 15.7 on EnterpriseOps-GYM after RL training
- Qwen3.5-4B: 45.9 -> 56.0 on AIME26 after training on AgentMercury
- World authoring success: 3.3% -> 83.3% on held-out scenarios after fine-tuning
Scenario-grounded, verifiable worlds can provide transferable learning signals beyond benchmark-specific environment design.