What's this about?
This paper argues that for training multimodal agents, *which* practice environments you choose matters more than simply piling on more of them — and it shows how to design an effective distribution of environments rather than scaling count blindly.
What exactly is an "agent environment"?
An environment is the simulated world an agent trains in — a website, a game, a desktop — where it takes actions and gets feedback. "Environment scaling" is the assumption that more environments = better agents. This work says diversity and design of those environments drive learning more than raw quantity.
Why is this notable?
It's a corrective to a popular scaling reflex, and practically useful: teams building agent training pipelines can spend on the right environments instead of just more of them.