Google Launches Antigravity, an Agent-First IDE Built for Gemini 3

Google has released Antigravity, an agentic development platform available in public preview today at no charge for individuals, with generous rate limits on Gemini 3 Pro usage.

The launch rides on Gemini 3, which Google calls its most intelligent model and "a step-change for agentic coding." The company frames Antigravity as the answer to a structural question: as agents run longer without human intervention, the interface between developer and agent has to change.

"We want Antigravity to be the home base for software development in the era of agents," Google writes in the announcement. "Our vision is to ultimately enable anyone with an idea to experience liftoff and build that idea into reality."

The move lands as the market for AI coding tools intensifies, with Google now competing directly against incumbent agentic IDEs and coding agents from rivals including Anthropic, OpenAI, and Microsoft-backed Cursor.

Two surfaces, one platform

Antigravity ships with a familiar "Editor view" — an AI-powered IDE with Tab completions, in-line Commands, and an agent running in the side panel. The more unusual addition is the "Manager" surface, described as "a mission control for spawning, orchestrating, and observing multiple agents across multiple workspaces in parallel."

The Manager inverts the current paradigm, according to Google: instead of agents being embedded within surfaces like an editor or browser, the surfaces are embedded into the agent. Google chose not to merge the synchronous Editor and asynchronous Manager into a single window, optimizing instead for fast handoffs between the two.

Google acknowledges the current ceiling: models can now run for extended periods across multiple surfaces without intervention, but "not yet for days at a time." The company says it is approaching a world where developers interact with agents at higher abstractions than individual prompts and tool calls.

Four design tenets

Google structures the product around trust, autonomy, feedback, and self-improvement.

On trust, Google argues existing tools fail users in one of two ways: they either show every raw tool call, or they show only the final code change with no way to verify how the agent got there. Antigravity instead groups tool calls into tasks and surfaces what Google calls Artifacts — "task lists, implementation plans, walkthroughs, screenshots, and browser recordings" — deliverables that are easier for a human to validate than raw logs. Google says it puts "a concerted emphasis for the agent to thoroughly think through verification of its work, not just the work itself."

On feedback, Antigravity supports Google-doc-style comments on text Artifacts and select-and-comment feedback on screenshots. Feedback is "automatically incorporated into the agent's execution without requiring you to stop the agent's process." Google's reasoning is pragmatic: "An agent being able to complete 80% of the work should be useful, but if there is no easy way to provide feedback, then it becomes more work than benefit to resolve the remaining 20%."

On self-improvement, agent actions both retrieve from and contribute to a knowledge base, capturing explicit information like useful code snippets and architecture, as well as the steps taken to complete particular subtasks.

Antigravity also includes browser control capabilities, letting agents autonomously execute end-to-end tasks across surfaces.

Multi-model, cross-platform, free

The preview runs on macOS, Linux, and Windows. Beyond Google's Gemini 3, the agent supports Anthropic's Claude Sonnet 4.5 and OpenAI's GPT-OSS, giving developers model optionality.

Rate limits refresh every five hours and are correlated with the amount of work the agent performs rather than the number of prompts. Simple tasks consume less budget; harder tasks consume more. Google's modeling suggests "a very small fraction of power users will ever hit the per-five-hour rate limit."

Google says new features will drop frequently, with updates posted to its blog and its X, LinkedIn, and YouTube channels. The company positions the current preview as a starting point for what it calls "the next fundamental step function in agent-assisted development" — a bet that as Gemini models continue to get smarter, asynchronous, agent-first tooling becomes the default way software gets built.