World models are the new eval sandbox for agents.

DeepMind's Genie 4 is a real-time interactive world model with native audio and natural-language steering. They pitch it for training and test environments for robots and agents, including Gemini models planning steps that Genie renders.

It's a research preview for trusted testers, coming later to Google AI Ultra (Project Genie) and early Gemini API / AI Studio partners. DeepMind is clear on limits: detail and object permanence can drift over longer sessions, interaction is on the order of a few minutes, and text rendering is still weak.

Worth saving: I'd treat a world model as an eval sandbox when the session is short and the failure modes are logged — not as a production test harness yet.

What would you actually test an agent against in a simulated world first?

Sources:
https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4/