Agentic RL data scaling laws reportedly rank environments above tasks, then trajectories

A user describes a paper on scaling laws for agentic reinforcement-learning data across tasks, environments and trajectories. They say environments rank above tasks, which rank above trajectories, across model sizes, families, domains, and both PPO and GRPO. The user says the environment advantage appears to come mainly from genuinely different worlds, while changing one world's structure helps about as much as adding tasks. They also say the paper's budget allocation recipe gives up to 1.7x compute efficiency over generating more trajectories.