Google launches Antigravity 2.0: Gemini 3.5 Flash engine, dynamic sub-agent, and scheduled scheduling features in place

Google announces the launch of Antigravity 2.0 on the first day of I/O 2026, a standalone desktop application supporting macOS, Linux, and Windows, completely separated from the original Antigravity IDE.
(Background: Google officially releases "Gemini 3"! Reaching the top as the world's smartest AI model, what are the highlights?)
(Additional context: Google launches the most powerful "autonomous agent and programming development" model Gemini 3.5 Flash! Writes an operating system in 12 hours at a cost of less than $1,000)

Table of Contents

Toggle

• Gemini 3.5 Flash engine: 4 times faster, upgraded agent capabilities
• Making agent behaviors controllable and automatable
• From tool to platform

On the first day of Google I/O 2026, Google completely separated the core agent interface from the Antigravity IDE, which was only released in November 2025, turning it into a brand-new desktop app named Antigravity 2.0.

In the official explanation, Google directly admits that bundling the IDE and agent together in one app "is not intuitive," especially for users unfamiliar with coding, creating a barrier. The solution in 2.0 is: remove the IDE, keep only the agent.

Gemini 3.5 Flash engine: 4 times faster, upgraded agent capabilities

Antigravity 2.0 is powered by Gemini 3.5 Flash. According to benchmarks released by Google, Gemini 3.5 Flash outperforms Gemini 3.1 Pro in coding and agent tasks, being 4 times faster.

One of the core mechanisms introduced in the new version is dynamic subagents. Simply put, the main agent can define and call subagents during execution to complete subtasks, preventing the main agent's context window from being overwhelmed, while allowing multiple subtasks to run in parallel.

This is paired with asynchronous task management: tasks can run in the background without blocking the main agent from processing other commands.

However, these features have already been implemented in Claude Code and Codex for some time, so they are not particularly new or exciting.

Making agent behaviors controllable and automatable

At the same time, version 2.0 introduces JSON hooks. Users can use a simple JSON configuration file to intervene, rewrite, or block the agent's behavior at critical moments during execution. Non-engineers can also perform fine-grained control over the agent without needing to understand underlying code.

Another new mechanism is Scheduled Tasks: define trigger times via cron scheduling, allowing the agent to automatically start at specified times. Paired with the new slash command /schedule, users can directly input commands like "整理收件匣摘要每天早上九點" ("Summarize inbox every morning at 9 AM").

New slash commands also include /goal (set a target and run until completion without interruption), /grill-me (agent proactively asks clarifying questions), and /browser (explicitly control whether to use a browser).

From tool to platform

The complete Antigravity platform now consists of five layers: desktop app, CLI, SDK, Gemini API Managed Agents, and Enterprise Gemini Platform. Antigravity also natively integrates with AI Studio, Firebase, and Android, with AI Studio projects capable of directly exporting with context.

Subscription pricing has been adjusted accordingly: a new AI Ultra $100/month plan (up to 5 times the usage limit for Pro), while the original AI Ultra has been reduced from $250 to $200/month (up to 20 times the usage limit for Pro).

Google has not immediately shut down Antigravity IDE; both apps will coexist temporarily. However, Google has announced that the Agent Manager will be removed from the IDE, returning the IDE to a purely agentic development environment.