At this year's Google I/O 2026, the most exciting technology announcement for developers was undoubtedly the upgrade of Google's "Antigravity" AI development environment, launched just six months prior, to version 2.0. This version officially transforms Antigravity from a simple "code editor" into an unabashedly agent-centric "Autonomous AI agent development and management platform." Through new desktop applications, CLI, and SDK, coupled with the Gemini 3.5 Flash model announced on the same day, developers can now act as "project directors," directly commanding multiple AI agents to perform parallel tasks. They can even create an operating system from scratch within 12 hours that can smoothly run the classic game Doom.
Three new interfaces are launched simultaneously, unleashing the potential of "sub-agent collaboration".
Antigravity 2.0 completely breaks away from the framework of a pure code environment, redesigning the interface around dialogue with agent AI, outputs, and multi-agent scheduling, and introducing three operating interfaces for different development habits:
• Standalone desktop applications:Transform Antigravity from a purely coding environment into an agent dispatch center that anyone can use. You can have one agent writing the front end, another generating the brand visuals, and yet another planning the product architecture.
• Lightweight CLI (Command Line Interface):Designed for heavy developers who prefer to work on the terminal, it allows for instant wake-up and activation of AI agents without the need for a graphical interface.
• Deeply optimized SDK:Open programmatic access capabilities allow developers to directly call the same proxy framework that Google uses to drive its own products.
Among the most anticipated features is the "Sub-agent teamwork" function, which is currently in the early research preview stage. This allows multiple AI sub-agents to work in parallel on different sub-tasks and then perfectly integrate the results to complete an end-to-end code pipeline.
Demonstration: A system was written in 12 hours for less than $1000.
To demonstrate the limits of Antigravity 2.0 combined with Gemini 3.5 Flash, Google presented a visually striking real-world test case during its keynote address: having a team of AI agents write an operating system from scratch.
This massive project lasted over 12 hours, with multiple sub-agents issuing more than 15000 model requests and processing up to 26 billion tokens. From the scheduler and memory management to the file system, every line of code was autonomously written, reviewed, and tested by AI, and the entire process cost less than $1000 in API call fees.
In the live demonstration, this operating system, built from scratch by AI, not only successfully ran the classic game "Doom," but also automatically wrote and integrated the driver program after receiving instructions from the presenter when it was found that the display and keyboard drivers were missing.
Powered by Gemini 3.5 Flash: 12x faster inference speed and extreme cost advantage
The biggest contributor to Antigravity 2.0's impressive performance is the Gemini 3.5 Flash model, which it co-developed with.
• The speed of inference is skyrocketing:The standard version of Gemini 3.5 Flash has a token output speed that is 4 times faster than that of a large language model of the same level; while the optimized version running in the Antigravity environment can be up to 12 times faster, perfectly solving the latency problem of multiple agents executing tasks for a long time.
• Enterprise-level cost killer:Google points out that if leading companies in the industry migrate 80% of their workloads, which amount to 1 trillion tokens per day, from other cutting-edge models to Gemini 3.5 Flash, they could save an astonishing $10 billion annually.
Code Bender debuts: A cybersecurity gatekeeper who transitions from programming to "bug hunting and patching".
With the exponential increase in AI-generated code, code security has become a major concern. Google has responded by releasing "Code Bender," a cybersecurity tool built using Gemini's advanced reasoning capabilities.
Unlike most scanning tools on the market that can only "find" vulnerabilities, Code Bender's biggest differentiating value lies in its ability to directly "generate patch code".
This technology is a direct descendant of CodeMender, which Google DeepMind released last year. Google has also confirmed that CodeMender's technology will be integrated into the Gemini Enterprise agent platform, which will automatically identify and assist in fixing major vulnerabilities in the enterprise development process through a powerful AI security agent.
Analysis: Overwhelm competitors with "infrastructure" and "extreme speed"
The release of Antigravity 2.0 shows that Google is no longer satisfied with just being a "code writing aid" like GitHub Copilot, Cursor, or Claude Code, but wants to create a "fully automated proxy factory" for the software development industry.
