The landscape of AI-assisted development has evolved rapidly, moving from simple code completion to fully integrated "agentic" environments. The latest entrant to this competitive space is Google Antigravity, a public preview release that promises to redefine how developers interact with their IDEs.
Antigravity offers a familiar VS Code-like interface but introduces a sophisticated "Agent Manager" designed to spawn, coordinate, and test autonomous coding tasks. At the heart of this system lies a diverse selection of large language models (LLMs).
The Power of Gemini 3
The core engine driving Google Antigravity is the Gemini 3 Pro model, which is available in two distinct configurations: "High" and "Low."
This tiered approach allows developers to balance computational cost and speed against reasoning depth, depending on the complexity of the task at hand.
Perhaps the most intriguing aspect of the Gemini 3 integration is its multimodal potential. While the current usage focuses on code and image context, the implication is that future iterations could allow developers to feed video context—such as a screen recording of a bug or a feature demo—directly into the agent to drive development.
Beyond Google: A Multi-Model Approach
One of Antigravity’s most surprising features is its willingness to step outside the Google ecosystem. The platform includes access to Claude Sonnet 4.5, widely regarded as one of the top-tier models for coding tasks. This inclusion suggests that Antigravity aims to be a model-agnostic platform where the best tool can be used for the job, rather than a walled garden for Google products.
However, the model selection also includes some curiosities. The platform lists "GPTO OSS 120," described as an open-source model from OpenAI.
Planning, Fast Mode, and Autonomous Testing
The choice of model heavily influences the two primary modes of operation: "Fast Mode" and "Planning Mode." In Planning Mode, the models generate a step-by-step roadmap before writing code, allowing the user to intervene, skip steps, or provide feedback on specific images or architectural decisions.
However, the true "killer feature" powered by these models is the autonomous testing capability. Unlike standard IDEs, Antigravity uses its agents to physically interact with the browser. It simulates mouse movements, clicks buttons (like "Shop Now"), scrolls, and hovers over elements to verify UI responsiveness. This level of semantic understanding—where the model reasons through the UX flow—sets a new standard for what developers can expect from an AI pair programmer.
While Google Antigravity is still in a rate-limited public preview, its integration of Gemini 3 Pro alongside Claude Sonnet 4.5 offers a glimpse into a future where IDEs are not just text editors, but command centers for intelligent, multimodal agents.
You can follow me on GitHub, where I'm creating cool projects.
I hope you enjoyed this article, until next time 👋
Top comments (0)