Why agents need diverse worlds, not just more tasks

This position paper identifies a critical gap in how we train generalizable agents: current scaling efforts add more experience and tasks within fixed rule-sets, leaving agents brittle when interfaces, dynamics, or feedback change. The authors propose environment scaling—systematically varying the executable rules agents interact with—as distinct from trajectory and task scaling. They offer a taxonomy distinguishing these three scaling types and survey construction paradigms (programmatic generators for control versus generative world models for breadth), then discuss coupling environment scaling with learned adaptation mechanisms. The work argues that scalable environments are foundational for building agents robust to world-level distribution shifts.