Google Rebuilds Antigravity 2.0 Around Gemini 3.5 Flash to Replace Coding With Orchestration
Google I/O 2026 reframes AI tooling as infrastructure: Gemini 3.5 Flash, Antigravity 2.0, WebMCP, and an Android migration agent signal a full shift to autonomous development workflows.
- 🚀 Gemini 3.5 Flash is now GA: 4x faster than rival frontier models, beats Gemini 3.1 Pro on most benchmarks, priced at $1.50/$9.00 per MTok (input/output).
- 🤖 Antigravity 2.0 ships as a full desktop app with parallel subagent orchestration, a new CLI replacing Gemini CLI (shutting down June 18), an SDK, and Managed Agents via a single Gemini API call.
- 🌐 WebMCP is a proposed W3C standard (co-developed with Microsoft) letting websites expose structured JS/HTML tools to browser AI agents — origin trial opens in Chrome 149, with 67% fewer errors vs. visual scraping.
- 📱 Android Migration Agent (preview) converts iOS, React Native, or web framework apps to native Kotlin in hours instead of weeks, using Jetpack Compose best practices.
- 🛠️ Chrome DevTools for Agents and Modern Web Guidance (100+ use cases) give coding agents automated quality audits and expert-vetted web skills, installable via
npx modern-web-guidance install. - ⚠️ Rocky launch: the 2.0 auto-update broke existing workflows, a sandbox escape vulnerability was found within 24 hours, and terminal sandboxing is macOS-only — no Linux or Windows equivalent.
Google used its I/O 2026 developer keynote to make a pointed argument: the era of AI-assisted coding is over. The era of AI-agentic coding has begun. The announcements were not a single model drop or a feature update , they were a coordinated stack, from model to IDE to browser standard, designed to move the unit of developer work from a prompt to an autonomous workflow.
The engine: Gemini 3.5 Flash
At I/O 2026, Google launched Gemini 3.5 Flash, combining frontier intelligence with speed. It outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster than other frontier models, providing the high-speed engine needed for real-world agentic workflows. That speed gap matters more than it sounds. In agent loops, every reasoning step, every tool call, every synthesis pass costs latency. A 4x speed advantage compounds across dozens of turns in a single task.
Gemini 3.5 Flash is priced at $1.50/MTok input and $9.00/MTok output. Compare that to Claude Opus 4.7 at $15/$75 per MTok , that's 10x cheaper on input. For coding agents that make many small calls , writing tests, fixing linting errors, generating boilerplate , Gemini 3.5 Flash's price point starts to look compelling.
The platform: Antigravity 2.0
The 2.0 release is a complete rebuild. Gone is the VS Code fork that defined Antigravity 1.0. In its place: a standalone desktop app for multi-agent orchestration, a Go-based CLI that replaces the Gemini CLI entirely, an SDK for self-hosted agent deployment, and Managed Agents in the Gemini API that give you an autonomous coding agent with a single API call.
The headline capability is parallel subagent orchestration , a concept worth unpacking. Rather than a single agent working through a task sequentially, you hand the main agent a complex task like "refactor our auth module, update the corresponding tests, and document the API changes." The primary agent breaks that task into components, spawns subagents for the test update and documentation pieces, and those subagents run asynchronously in isolated contexts. The main conversation thread does not stall while subagents work. You get notified when they complete, and their outputs come back into your primary thread as structured results.
Here is a conceptual look at what orchestrating those subagents via the SDK looks like:
import { AgentOrchestrator } from '@google/antigravity-sdk';
const orchestrator = new AgentOrchestrator({
model: 'gemini-3.5-flash',
parallelAgents: 4,
sandboxed: true,
});
const result = await orchestrator.run({
intent: "Add async payment processing with OpenTelemetry tracing",
subagents: [
{ role: 'refactor', focus: 'async patterns' },
{ role: 'observability', focus: 'tracing instrumentation' },
{ role: 'testing', focus: 'integration test suite' },
{ role: 'review', focus: 'cross-agent consistency' }
]
});Google's Agent-to-Agent (A2A) protocol, backed by over 150 organizations, is natively supported. Where MCP handles agent-to-tool calls, A2A handles agent-to-agent delegation: one specialist agent handing off a subtask to another. Combined with the parallel subagent architecture, Antigravity 2.0 has a multi-agent coordination story that no current competitor matches out of the box.
The full Antigravity 2.0 surface now spans five entry points:
- Desktop app , standalone, multi-agent orchestration with a Mission Control dashboard for managing parallel agents
- Antigravity CLI (
agy) , Go-based terminal binary with the same agent harness as the desktop, built for CI/CD pipelines and power users - Antigravity SDK , programmatic control over the harness, so a development organization can deploy the agent loop on its own infrastructure rather than depending on Google's hosted runtime
- Managed Agents in the Gemini API , a single API call spins up an agent that reasons, calls tools, and executes code inside an isolated Linux environment; sessions are stateful and resumable, so files and context persist across turns
- Gemini Enterprise Agent Platform , Antigravity 2.0 and the Antigravity CLI available to Google Cloud customers, connecting directly to Google Cloud projects
Pricing: a new $100/month AI Ultra plan brings a 5x higher usage limit in Google Antigravity than the Google AI Pro plan. The Pro tier sits at $20/month, directly competing with Cursor Pro and Windsurf Pro.
The web layer: WebMCP
Alongside the agent platform, Google proposed a new browser standard that could quietly reshape how AI agents interact with the entire web. Google unveiled WebMCP at I/O 2026, a proposed W3C standard co-developed with Microsoft that gives websites a structured way to expose callable actions directly to browser-based AI agents running inside Chrome.
The problem it solves is fundamental. If you've ever watched an AI agent "use" a website, you know the pattern: taking screenshots of pages, guessing which blue rectangle is the "Submit" button, scraping DOM elements and hoping nothing changed, clicking around until something works. This is billion-parameter models pretending to be humans, pixel by pixel.
Instead of agents scraping screens or reverse-engineering page layouts, developers declare named JavaScript functions and HTML forms as machine-readable endpoints , "search inventory," "initiate checkout," "submit support request." Agents invoke these with explicit user authorization, making the interaction auditable rather than opaque. The performance improvement is significant: structured WebMCP calls produce 67% fewer errors and 45% better task completion rates compared to visual scraping.
The experimental WebMCP origin trial starts in Chrome 149, with support for Gemini in Chrome coming soon. Firefox support is committed for Q3 2026; Safari is expected Q4. The spec is being pitched to W3C, meaning adoption could extend beyond Chrome to other browsers if the spec lands.
The Android angle: migration in hours, not weeks
One of the most practically significant announcements for mobile teams was the Android Migration Agent. The Migration Assistant in Android Studio is designed to port apps from platforms like iOS, React Native, or web frameworks to native Android. By simply selecting an existing project, developers can have the agent intelligently map features, convert assets like storyboards and SVGs, and implement Android best practices using Jetpack Compose and recommended Jetpack libraries , effectively transforming what used to be weeks of manual porting into a streamlined agentic workflow that only takes hours.
It's available in preview and will ship in a future Android Studio version. Separately, developers can now use prompts to build native Android apps in Google AI Studio using Kotlin, Jetpack Compose, the latest APIs, and Google's recommended developer patterns , prototype an app, test it in an embedded emulator, and deploy it to physical devices.
Chrome DevTools for agents and Modern Web Guidance
Two supporting tools round out the web story. Chrome DevTools for Agents extends the familiar DevTools interface to autonomous workflows: it lets AI automate quality audits and debug in real time without manual oversight. Modern Web Guidance is a curated skill pack for coding agents , a set of evergreen, expert-vetted skills that guide coding agents across common use cases to build web experiences that are more accessible, performant, and secure. It integrates directly with Baseline, letting you focus on what you want to build while your tools automatically figure out the right features and fallbacks. Install with a single click in Google Antigravity, through npx, or as an extension in a coding agent.
The rough edges
The launch was not without friction. The 2.0 auto-update removed the built-in code editor entirely, breaking active production workflows for developers who relied on the IDE. The credit system switch confused users. And a prompt injection vulnerability leading to sandbox escape was discovered within 24 hours of public availability.
The sandboxing story also has gaps. Agent-executed terminal commands run in a kernel-level isolated environment that restricts writes to your project folder, using Apple Seatbelt , the same technology macOS uses to sandbox App Store applications. But this is macOS only. Linux and Windows teams get no equivalent, which is a significant gap for CI/CD pipelines and server-side development environments.
The framing from Augment Code is: "You are the orchestrator, not the developer." They recommend using Antigravity alongside Cursor or Claude Code rather than as a replacement. Claude Code is described by reviewers as terminal-first with stronger reasoning for complex backend and architectural work. The general view is that Antigravity's integration depth with Google's broader stack , Firebase, AI Studio, Android , gives it an advantage for teams already invested in that ecosystem, while Cursor's polish and Devin's collaboration model maintain their appeal for teams that are not.
What this means for how software gets built
The real shift at I/O 2026 was not any single product. Google used the developer keynote to reframe the agentic era as infrastructure, not a model release. Gemini 3.5 Flash is the engine. Antigravity 2.0 is the runtime. WebMCP is the browser interface. The Android Migration Agent is the first proof that the stack can eat a real, painful developer workflow whole.
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. Whether that is a productivity unlock or a new class of debugging problem is the question the rest of 2026 will answer.
All 85+ sessions from I/O 2026 are available on demand at io.google/2026. The Antigravity platform and full developer keynote recap are live now. The WebMCP origin trial opens in Chrome 149.