Antigravity 2.0 Decouples Agent Environments With Gemini 3.5

Google DeepMind announced Google Antigravity 2.0 at Google I/O, restructuring its experimental agent IDE into a decoupled orchestration platform. Driven by the newly integrated Gemini 3.5 Flash model, the release moves agent management out of the code editor and into dedicated desktop, terminal, and SDK environments.

Platform Architecture and Surfaces

The original VS Code-based environment has been broken into four distinct components. The core application is now the Antigravity 2.0 standalone desktop client for macOS, Linux, and Windows, functioning as a mission control surface for orchestrating background tasks. To accommodate this shift, the IDE extension will eventually lose its Agent Manager view, restricting its scope to standard code editing tasks.

Developers handling high-speed terminal execution can now use the new Antigravity CLI, written entirely in Go. This tool officially deprecates the existing Gemini CLI, which Google will sunset on June 18, 2026. Custom agent deployments are managed through the Antigravity SDK, available via pip install google-antigravity, which provides boilerplate integration in under 20 lines of Python.

Agentic Orchestration and Subagents

Antigravity 2.0 introduces Dynamic Subagents to manage context saturation during extensive tasks. A primary agent can programmatically spawn child processes to handle parallel execution streams like backend development or security audits. In a keynote demonstration, Google showed the runtime orchestrating 93 parallel subagents to build an operating system within 12 hours. This parallel architecture shifts how to implement multi-agent coordination patterns in production environments.

The orchestration layer includes a built-in Chromium-based Browser Agent for visual QA loops and web interaction. Developers configure these roles and define specialized agent skills using the new AGENTS.md and SKILL.md file standards. Agent behavior can be intercepted and modified during runtime using standardized JSON hooks, while the /schedule command allows for cron-like recurring execution.

Gemini 3.5 Flash Performance

The platform relies entirely on Gemini 3.5 Flash, prioritizing raw execution speed for high-volume automation. Google records the model’s output at 289 tokens per second, achieving a 12x speed improvement over previous frontier models for agentic tasks.

Metric / Benchmark	Antigravity 2.0 Result
Underlying Model	Gemini 3.5 Flash
Output Speed	289 tokens per second
SWE-bench Verified	76.2%
Orchestration Scale	93 parallel subagents demonstrated

The 76.2% score on SWE-bench Verified places the environment slightly behind Anthropic’s Claude Sonnet 4.5. The speed tradeoff positions the platform for highly parallel workflows where context windows are distributed across multiple isolated agents rather than stacked in a single monolithic prompt.

Enterprise API and Availability

For production deployments, Google introduced the Managed Agents API. A single API call spins up a fully isolated, stateful Linux environment designed for multi-turn sessions. Access requires the new AI Ultra tier, priced at $100 per month, which increases usage limits by five times over the standard AI Pro plan. Individual developers retain access to a free tier with lower throughput caps during the public preview.

If you currently depend on the Gemini CLI for automated workflows, you must migrate your scripts to the new Go-based Antigravity CLI before the June 18 deprecation date. Transition your prompt structures into the new AGENTS.md and SKILL.md configuration formats to ensure your routines run correctly under the new execution engine.