Google's Antigravity 2.0 Rebuilds Software Development Around 93 Simultaneous AI Agents

Google just made its biggest bet yet that the future of software development is not writing code , it's directing agents that write it for you. At Google I/O 2026, the company unveiled Antigravity 2.0, a standalone desktop application rebuilt from the ground up around multi-agent orchestration. This is not a feature refresh. Antigravity 2.0 is a full platform pivot from AI-assisted coding to multi-agent orchestration as the core development model.

From one IDE to a whole stack

The original Antigravity that launched in November 2025 was a single editor. The 2.0 release is a full agent-first development stack. It is now a desktop IDE, a CLI, an SDK, a Managed Agents tier inside the Gemini API, and an enterprise deployment path through the Gemini Enterprise Agent Platform. The thread Google posted announcing 2.0 features racked up over 254,000 views and 3,350 likes, signaling the developer community is paying close attention.

When Google launched the original Antigravity IDE in November 2025, there was no agent-first GUI surface in the market. They wanted to prove that such a surface worked, at least for software development. So while the core was a familiar agent-powered IDE, they introduced the Agent Manager , a second surface that stripped away much of the IDE UI, allowing users to focus on agent conversations, artifacts, and multi-agent management. With 2.0, that Agent Manager is now its own standalone product entirely.

Five features that change the loop

The 2.0 announcement thread broke down five new capabilities. Here is what each one actually means in practice:

• Dynamic Subagents: The main agent can define and spawn specialized subagents on the fly to tackle focused subtasks in parallel, without polluting its own context window. This addresses a fundamental limitation of single-agent systems , context window overload. When facing complex tasks, your main AI agent can now automatically spin up specialized helper agents to handle different aspects of the work simultaneously, mirroring how human teams delegate tasks to specialists.
• Asynchronous Task Management: Long-running operations run in the background, so neither the main agent nor the UI is blocked. You can kick off a multi-hour refactor and keep using the interface for other tasks while it runs.
• Scheduled Tasks: Users can define cron-style schedules that trigger agents automatically, including daily digests, hourly health checks, and monthly reports, without manual invocation. This turns the agent from a tool you prompt manually into something closer to an automation pipeline with a conversational interface.
• JSON Hooks: You can now define hooks in a simple JSON format, allowing you to intercept and control the Antigravity agent's behavior. The policy system lets you lock things down declaratively , denying everything by default, allowing specific tools like file reads, and requiring human approval for shell commands.
• Voice: Google is adding native voice command support to Antigravity, just like it has added to multiple consumer products, including Gmail and Docs. Real-time transcription means you can speak directly to your agents.

The engine underneath: Gemini 3.5 Flash

Underpinning the entire ecosystem is Gemini 3.5 Flash, which Google is setting as the default model across Antigravity. According to the Google team, 3.5 Flash outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster than other frontier models , a speed advantage that is practically significant when multiple agents are running in parallel, since model latency compounds across concurrent agent calls.

A lot of this is powered by the company's new Gemini 3.5 Flash model, which was co-developed using Antigravity, Google said. The platform also supports Anthropic's Claude Sonnet 4.5 and OpenAI models, but the platform is optimized for Gemini, which means lower cost and lower latency on Google's family.

The I/O demo that made jaws drop

The headline moment from the I/O keynote was not a slide deck. Varun Mohan showed the system building a complete operating system from scratch in 12 hours for under $1,000. The demo ran 93 subagents simultaneously, processing 2.6 billion tokens across 15,000 model requests. Mohan then ran Doom on the new OS live on stage. Whether that maps to your daily standup is a fair question, but the underlying infrastructure it demonstrated is real.

Early hands-on reports from developers are largely positive on the subagent model itself. One developer went in expecting a single AI thread pretending to be multi-agent, and got something genuinely different , you can watch the subagents in the Antigravity sidebar, each one with its own task, status, and output log. Another reported that the agent did not forget requirements across long prompts, navigated files with precision, and auto-ran build commands , features that normally took a full day were completed in under an hour.

The rollout that made developers furious

The launch was not without friction. Google pushed an automatic update to Antigravity on May 19 that gutted the code editor. Developers opened their computers to find terminals, file explorers, and editing tools gone. Google announced Antigravity 2.0 at I/O the same day, but the automatic rollout turned functional development environments into chat interfaces.

Subreddits and Google AI Developer forums filled with complaints within hours. One developer said it "reeks of non-technical people shipping code to production" after active projects became unusable. Another described the agent-first pivot as a massive step backward. Google split Antigravity into three separate downloads. The automatic update installed Antigravity 2.0, which does AI agent work. Developers who want their code editor back need to download Antigravity IDE separately.

There is also a hard deadline for Gemini CLI users. The Antigravity CLI fully replaces the Gemini CLI, which is getting retired. There is an end date of June 18, 2026 for AI Pro, AI Ultra, and free-tier users. This is not a gentle deprecation , it is a hard cutoff.

The SDK and Managed Agents API: the quiet power move

Beyond the desktop app, the SDK gives you the same agent runtime that powers Antigravity 2.0 and the Antigravity CLI. Your agent inherits a capable execution environment , a rich built-in toolset, a declarative safety-policy engine, lifecycle hooks for observing and steering every tool call, and stateful multi-turn sessions that persist across interactions.

With a single API call, developers can spin up an agent that reasons, uses tools, and executes code in an isolated Linux environment. Managed Agents are powered by the Antigravity agent harness, built on Gemini 3.5 Flash, and available via the Interactions API and in Google AI Studio. Each interaction creates an environment that can be resumed in follow-up calls with all files and state intact, enabling seamless multi-turn sessions without reinitializing context.

Here is a minimal Python example of what spinning up a Managed Agent looks like with the SDK:

from google.antigravity import Agent
from google.antigravity.hooks.policy import deny, allow, ask_user
policies = [
    deny("*"),           # Block everything by default
    allow("view_file"),  # Except reading files
    ask_user("run_command", handler=my_handler),  # Human approval for shell
]
async with Agent(policies=policies) as agent:
    result = await agent.run("Refactor the rate-limit middleware in ./services/api")
    print(result.structured_output())

What it costs and what you get

Google Antigravity is available in public preview, at no cost for individuals. The cross-platform solution is compatible with macOS, Windows, and Linux, offering model optionality with generous rate limits on Gemini 3 Pro, and full support for Anthropic's Claude Sonnet 4.5 and OpenAI's models.

The mental model shift that matters most

Antigravity 2.0 introduces a structurally different loop: the developer defines a goal and constraints, the agent spawns specialized subagents, parallel execution happens across tasks, and the developer evaluates outputs and refines direction. The developer is no longer spending primary effort on producing implementation details , they spend primary effort on defining objectives, setting constraints, and evaluating outcomes. The center of gravity moves from writing toward orchestrating.

JSON hooks and project-level scoping are the underappreciated features here , they are what make enterprise adoption credible. For organizations considering agent adoption, those controls matter as much as raw capability. The problem is rarely whether an agent can do something once; it is whether the agent can be constrained, reviewed, and repeated safely.

The biggest risk the community is flagging is not hallucination , it is developer overconfidence in outputs they did not produce and did not fully verify. The infrastructure is real. The judgment about when to trust it still belongs to you.