Google’s Antigravity: A New AI Tool for Coding That Puts Agents in Charge
Google just dropped Antigravity, a fresh coding tool that flips the script on how developers work with AI. It pairs with the company’s new Gemini 3 model and lets AI agents take the wheel on tasks like editing code, running terminals, and even controlling browsers. No more babysitting chatty sidekicks—this thing aims to handle the heavy lifting so you can focus on the big picture.
What Antigravity Does
Picture this: instead of AI tucked into your editor like a helpful intern, Antigravity makes agents the stars of the show. They juggle multiple tasks across workspaces, plan steps, and execute them without constant hand-holding. As The Verge reports, it spits out “Artifacts”—think task lists, screenshots, and browser recordings—to prove what it’s done and what’s next. That way, you can verify the AI’s moves without drowning in logs.
The tool offers two views: a standard Editor mode that feels like your usual IDE, with agents chatting in a sidebar, and a Manager view for overseeing a squad of agents at once. Google calls it “mission control” for parallel workflows, which sounds about right if you’re wrangling complex projects.
Feedback loops are smart too. Leave comments on those Artifacts, and the agents adjust on the fly. They even remember past tricks, like code snippets or task steps, to get sharper over time. InfoWorld points out that developers act more like architects here, guiding agents through end-to-end jobs from research to full apps.
Tied to Gemini 3’s Brainpower
Antigravity runs on Gemini 3 Pro, Google’s latest AI upgrade announced alongside it. This model crushes benchmarks in coding and math—scoring 76.2% on SWE-bench for code generation and topping leaderboards like LMSYS Arena with 1,501 ELO points. Ars Technica notes it handles visuals better, reasons through PhD-level questions at 37.5% accuracy, and cuts down on hallucinations with a 72.1% score on fact-checking tests.
- Improved multimodal smarts for text, images, and video.
- Stronger agentic workflows, perfect for zero-shot tasks.
- Pricing starts at $2 per million input tokens in preview via Google AI Studio.
It’s not just hype; testers say Gemini 3 edges out rivals in real-world coding vibes. Antigravity taps into that for browser control and async interactions, evolving the IDE into something agent-first.
Who’s It For and How to Try It
Available now in free public preview for Windows, macOS, and Linux. It supports Gemini 3 Pro with generous limits (refresh every five hours), plus models like Claude 3.5 Sonnet and GPT-4o. Only heavy users might hit caps, Google says. Download it and see if it lives up to the “agent-first future” pitch—could save you hours on grunt work, or just add another layer of AI weirdness to your day.
