Google built AI agents a world to practice in. The physics still aren’t quite right.

Developers now possess a new arena for honing AI agents, thanks to Google DeepMind's introduction of Genie 4 on Tuesday. This world model creates interactive environments in real time, complete with simulated physics, native audio, and human motion.

While the photorealistic demonstrations are highly impressive, the key appeal for those building AI agents is the ability to run multi-step tasks within this simulated space before official evaluation.

Tejas Kulkarni, a research director at Google DeepMind, framed this approach at a press briefing as a step toward "these hybrid world models." In this setup, Gemini handles the code and text, while Genie translates that into a functioning scene.

This provides developers with the capacity to generate necessary test environments on demand. However, the consistency of these simulations mirroring actual world conditions remains an unresolved question regarding reliability.

Genie 4 leverages several Gemini models to maintain its simulations. Gemini Omni helps structure the environment, while Gemini 3.8 Flash tracks actions, plans subsequent moves, and adapts when changes occur.

Instead of rendering a final video, users can actively interact with the generated world, moving through it via keyboard, adjusting camera angles, or modifying the scene using text prompts without restarting the simulation.

A clear demonstration involves Apptronik's Apollo humanoid playing checkers, with Gemini 4 Argon planning each move and Genie 4 rendering the outcome as captions guide the scene. This supports Gemini Robotics research by allowing agents to be tested in dynamic settings without rebuilding every simulation.

A significant hurdle is that Genie 4 cannot always maintain environmental consistency; objects may shift or act unpredictably over extended sessions, and Google concedes the physics is not always accurate. This complicates agent evaluations, especially when tasks require memory of locations or returning to prior states.

Currently, only trusted testers can access Genie 4. Google intends to broaden access via Project Genie for Google AI Ultra subscribers and through the Gemini API and Google AI Studio for select partners, though pricing and wider release details are pending.