Antigravity Is Google’s New Agentic Development Platform

Google coming launched Antigravity, its latest research successful building an agentic improvement platform. Powered by nan caller Gemini 3 model, Antigravity combines what has go a beautiful modular AI-centric IDE pinch a fewer innovations that, according to Google, germinate nan IDE “towards an agent-first future.”

Antigravity will beryllium disposable for free successful nationalist preview for Mac, Windows and Linux starting today, pinch what Google calls “generous complaint limits” for Gemini 3 Pro usage.

Model Choice: Gemini, Claude and GPT-OSS

One absorbing facet present is that Gemini 3 isn’t nan only exemplary that developers tin choose. They will besides beryllium capable to usage Anthropic’s Claude Sonnet 4.5 and OpenAI’s open-weight GPT-OSS models to powerfulness nan agents.

“We will supply entree to models to nan grade we person capacity, pinch complaint limits to forestall maltreatment and that are refreshed each 5 hours,” Google explains.

Google and Anthropic precocious announced a awesome unreality partnership worthy billions of dollars. Maybe this was portion of nan deal.

Next-Gen Agentic Coding

The IDE itself will springiness developers entree to nan codification successful its editor position and each of nan accustomed trappings of immoderate modern improvement environment, including tab completions and in-line commands.

“We realized that LLMs — connection models — person really fundamentally changed really group codification and really we build software, really we bring ideas to life,” Koray Kavukcuoglu, nan CTO of DeepMind and Google’s main AI architect, said successful a property briefing up of today’s announcement. “So today, to push nan frontiers of really nan exemplary and nan IDE tin activity together, we are introducing Google Antigravity. It is our caller agentic improvement platform. I deliberation it is an important conception to deliberation astir really we create package pinch a dedicated supplier interface that enables developers to run astatine these higher, task-oriented levels.”

Agentic coding is evidently neither caller anymore nor anyplace adjacent nan highest of its hype cycle. But what we are seeing correct now is simply a changeless watercourse of innovations astir this concept, and while it’s intolerable to judge really good Google’s caller models execute successful this environment, Google is pushing nan letter cover astir nan wide developer experience.

Trust but Verify

The institution argues, for example, that akin devices successful this class thin to either show nan personification each action and instrumentality telephone nan coding supplier makes, aliases only show nan last codification without context. “Neither engenders personification spot successful nan activity that nan supplier undertook,” nan Antigravity squad writes successful today’s announcement.

In Antigravity, nan agents nutrient what Google calls “artifacts,” which it defines arsenic “deliverables successful formats that are easier for users to validate than earthy instrumentality calls,” including task lists, implementation plans, screenshots and browser recordings. The thought here, nan institution says, is to guarantee nan personification tin spot that nan supplier has verified its work.

“You tin look astatine really nan supplier has been executing, which states person been completed on nan way. The supplier will verify its activity done moving nan exertion wrong nan Chrome browser, and will, erstwhile it is finished, coming a walkthrough of really nan last merchandise is working,” Kavukcuoglu explained.

Agent Manager

Another illustration of really Antigravity is trying to span existing agentic coding paradigms and caller ideas is its attraction connected what it calls an “agent-first Manager surface,” which nan squad describes arsenic a benignant of Mission Control for spawning, orchestrating and watching agents arsenic they activity asynchronously connected a task.

Users who for illustration a much accepted model, pinch nan editor position and an supplier sheet connected nan side, tin take to do so, but nan Antigravity squad argues that “we are transitioning to an era, pinch models for illustration Gemini 3, erstwhile agents tin run crossed each of these surfaces simultaneously and autonomously.

“In summation to nan IDE-like Editor surface, we are introducing an agent-first Manager surface, which flips nan paradigm of agents being embedded wrong surfaces to 1 wherever nan surfaces are embedded into nan agent,” nan squad writes.

Steering nan Agent

To steer nan agent, developers tin constitute Google Docs-style comments connected nan artifacts nan supplier creates aliases springiness feedback connected nan screenshots it presents. Over time, nan supplier will study from this feedback by keeping an soul knowledge guidelines based connected erstwhile work. This whitethorn see snippets of codification aliases task lists of really it antecedently completed a task successfully.

What About Gemini CLI and Jules?

This caller instrumentality joins existing agentic coding devices connected Google’s platform, for illustration Gemini CLI, nan Jules coding agent and nan recently launched vibe coding mode successful Google’s AI Studio. They each attack this exertion from a different constituent of view, and some, for illustration Jules and now Antigravity, are explicitly meant to beryllium experiments. As of now, it seems Gemini CLI is nan work pinch nan astir traction. Whether immoderate of these different devices past successful nan agelong tally remains to beryllium seen.

Group Created pinch Sketch.