Antigravity 2.0 Updates: The 10 Changes Explained

If you use Google Antigravity and worry about agents running off and doing the wrong thing, the Antigravity 2.0 updates address exactly that. The new /plan command makes the agent write its steps first, and you approve them before any code runs.

What Did Google Ship in the Antigravity 2.0 Updates?

Google Antigravity, Google's agentic development platform, received ten updates between September and early October 2026, described in its latest mission report. The list covers a plugin store, a /plan mode, direct subagent messaging, the Gemini 3.8 Flash model, an updated agent harness, Windows sandboxing, and local model support in the SDK. The evidence for the specifics below is the mission report itself, so treat the performance figures as vendor-reported rather than independently measured.

The platform's own framing has shifted. What began as a coding environment with AI assistance now behaves more like a small operating system for agents: installable extensions, planned and approved work, isolated execution, and swappable models. The product page at Google Antigravity is the canonical reference for the current state.

The ten items, in the order the report presents them:

How Does the Antigravity Plugin Store Work?

Antigravity now has a Customizations tab that works as a plugin store: you browse, install, and manage plugins in one place. Google organizes it into curated shelves, some aimed at developers and some at plugins that connect to non-coding workflows.

Before this update, users found add-ons by searching the web and following manual setup steps. The store reduces that to a tab and a few clicks. Google describes the shelves as curated, meaning someone reviewed what appears on them, but the same basic caution applies as with phone apps: check what a plugin can access before you install it.

What Does /plan Mode Do for Agent Control?

Typing /plan makes the agent write out its intended steps before doing anything, and you can edit or delete steps before approving the work. This targets the most common failure with AI agents: they act immediately, build the wrong thing, and you spend time undoing it.

Google provides three review settings. You can require a plan review every time, let the agent review automatically, or skip review for small jobs. That choice determines how much control you keep. For multi-step tasks on important files, requiring review every time is the safe default.

How Do Direct Subagent Messages Change Team Workflows?

You can now message a subagent directly in the composer by typing @ and picking the subagent, instead of routing everything through the main agent. Subagents are helper agents that handle parts of a job: one researches, one writes, one reviews the result.

The practical gain is speed of correction. If a writing subagent drifts in the wrong direction, you can tap that one agent and say exactly what to change, without stopping the whole run. Previously, messages passed through the main agent, and details could get lost along the way, the way they do in a game of telephone.

What Is Gemini 3.8 Flash and the New Agent Harness?

Antigravity now runs on Gemini 3.8 Flash, Google's newest flash-class model, which Google says is built for long, tool-heavy agent jobs. Flash models trade some capability for speed. A model designed for long jobs keeps track of many sequential steps, such as opening a file, changing it, checking it, and moving to the next, without losing context. See Gemini for the model family.

Alongside the model, Google shipped an updated agent harness called Antigravity Agent 09-2026. The harness is the layer that tells the model how to act and how to use tools; the model is the engine, and the harness is the steering wheel. The same harness also appears in Google AI Studio and the Interactions API, so it is not locked to one product.

Google reports two numbers from its own internal testing: 40 percent fewer output tokens when the agent changes files, and up to 6 percent better task completion. These are vendor-reported figures from Google's setup, not independent benchmarks, and the exact workloads behind them are not published in the report.

Why Does Windows Sandboxing Matter for Antigravity?

Antigravity now supports file and network sandboxing on Windows, which limits the agent to the files and network destinations you allow. A sandbox is a safe box: if the agent makes a mistake, the mistake stays contained instead of touching the rest of the system.

Two details make this update more significant than it looks. First, the command-line sandbox on Windows no longer requires administrator rights, the highest permission level on a computer that company laptops usually block. Second, it works with Python, Git, and NPM out of the box. That means users on locked-down work machines can run the safe configuration without asking an IT department for elevated access.

Can You Run Antigravity Agents Offline With Local Models?

The Antigravity SDK now supports running agent workflows offline with models that live on your own machine. The SDK, or software development kit, is the toolset for developers building their own applications on top of the platform. Google named Gemma models, specifically a Gemma 4 26B-A4B variant served through LLM servers, plus OpenAI-compatible servers such as Ollama and vLLM.

This changes three things for builders. Privacy: your files never leave the computer. Reliability: no internet connection is required. Choice: you pick the model rather than relying only on Google's hosted ones. The tradeoff is that local models are usually weaker than large hosted ones, so this fits workflows where privacy or offline operation matters more than peak capability.

Note that this option lives in the SDK for developers. Running the standard Antigravity IDE experience fully offline is a different claim, and the report does not say that the main product works without Google's hosted models.

Do These Updates Make Antigravity an Agent Operating System?

The ten changes together give Antigravity the shape of an operating system for agents, though that is an editorial comparison rather than a Google label. The mapping is direct: the plugin store acts as the app store, the sandbox is the isolated space for each program, live subagent cards work like a task manager, and the ability to swap models resembles choosing hardware.

Any one update on its own would be a routine improvement. The combination of installation, planning, containment, and model choice in a single product is what makes this release feel like a category change. Whether it becomes a standard way of working depends on adoption, and the video's claim that most business owners have not caught up yet is the speaker's opinion, not measured data.

Frequently Asked Questions

• What are the biggest Antigravity 2.0 updates? The headline changes are the plugin store, /plan mode for reviewing agent steps before execution, direct @-messaging to subagents, the Gemini 3.8 Flash model, Windows file and network sandboxing, and local model support in the SDK.

• Is Gemini 3.8 Flash faster than previous models? Google describes it as a flash-class model built for long, tool-heavy agent jobs and reports 40 percent fewer output tokens on file changes in its own testing. These are vendor-reported figures; no independent benchmark is cited in the mission report.

• Can Antigravity run fully offline? Not the standard hosted experience. The offline capability lives in the SDK, where developers can point agent workflows at local models such as Gemma via Ollama or vLLM-compatible servers.

• Do I need administrator rights to use the Windows sandbox? No. The command-line sandbox on Windows no longer requires administrator rights and works with Python, Git, and NPM immediately, which makes it usable on locked-down company laptops.

• Are Antigravity plugins safe to install? Google curates the shelves in the plugin store, which adds a review layer. You should still check what a plugin can access before installing, the same habit you apply to phone apps.

Source video