What Are Antigravity 2.0 Updates?
Antigravity 2.0 updates are the ten major changes Google included in its latest antigravity mission report, covering September to early October. These updates turn antigravity from a simple AI coding tool into a full platform for building and managing AI agents.
In this episode, Julian Goldie's digital avatar walks through each update in plain language. The list includes a plugin store, a plan mode, direct sub-agent messaging, a new flash model, and even the ability to run agents offline with local models. Together, these changes make antigravity feel more like an operating system for AI agents.
Plugin Store: The New Customizations Tab
Antigravity now has a tab called Customizations. This is where you find plugins. You can browse them, install them, and manage them all in one place. Think of plugins as apps for your phone. The built-in tool does a few things, but plugins add many more.
Google set up the store with shelves. Some shelves hold tools for developers. Other shelves contain plugins that connect to places you already work, so you don't need to be a coder to find something useful. Before this update, you had to search the web and manually set up add-ons. Now it's one tab and a few clicks.
Are plugins safe? Google calls the shelves curated. That means someone picked what goes on them. You should still check what a plugin asks to access before installing it. It's the same habit you use with phone apps.
Plan Mode: Review Before the Agent Acts
The update Julian is most excited about fixes a common agent problem. When you type "/plan", the agent doesn't start working right away. Instead, it writes a plan first. It tells you the steps it wants to take, and you can edit that plan before anything happens.
Why does this matter? Because the biggest problem with AI agents is that they run off and do the wrong thing. You come back later to find it built something you never asked for. With plan mode, you catch that early. You read the plan, change a step, delete a step, and then say go.
Google gives you three ways to use plan mode:
- Require a review every time.
- Let it review automatically (or not).
- Skip review for small jobs.
You pick how much control you want.
Direct Messages to Sub-Agents
Antigravity's main agent can hand jobs to helpers called sub-agents. One might do research, one might write, and one might check the work. Before this update, you could only talk to the main agent, and it passed your message along. Like a game of telephone, things got lost.
Now you can message a sub-agent directly. In the composer where you type, use the @ sign. Pick the sub-agent and write to it. No middleman. This is useful when a sub-agent is writing something and takes a wrong turn. You don't need to stop everything. You just tap that one sub-agent and say "change this part."
This update also includes Livecards that show what each sub-agent is doing, so you can see their progress at a glance.
Gemini 3.8 Flash and the New Agent Harness
Antigravity now includes Gemini 3.8 Flash. Flash models are the fast ones. Google says this one is built for long jobs where an agent works for a while and for using tools. Some jobs take many steps, like opening a file, reading it, changing it, and checking it. This model keeps track of all those steps without getting lost.
There's also an updated agent harness called anti-gravity agent 09-2026. A harness is the part that tells the model how to act and how to use tools. Think of the model as the engine, and the harness as the steering wheel and pedals. This new version also shows up in AI Studio and the Interactions API, so the same agent style works in more places.
Google shared two numbers from internal testing:
- 40% fewer output tokens when the agent changes files.
- Up to 6% better task completion.
Windows Sandboxing and Local AI Models
The last two updates are easy to miss. First, antigravity now has file and network sandboxing on Windows. A sandbox is a safe box. The agent works inside it and can only touch the files and network places you allow. If it makes a mistake, the damage stays contained.
Google also made the command line tools in the Windows sandbox work without administrator rights. That means Python, Git, and NPM work right away on company computers. Previously, tools that needed admin rights were blocked for many workers. Now you can use them without asking IT for access.
The second missed update is local AI models in the SDK. Developers can run agent workflows offline, with no internet, using a model that lives on their own computer. Google mentioned options like Gemma 4, the 26B A4B version, through LiteLLM, and servers like Ollama and vLLM. This means nothing leaves your computer, which gives you privacy, reliability, and choice.
Plan Mode Setup: How to Get Started
Getting started with the new antigravity updates is straightforward. Here's how to begin with plan mode and other features:
- Open antigravity and go to the Customizations tab to browse the plugin store.
- Install a plugin that fits your work, like a tool for developers or a connection to a place you already use.
- In the composer, type "/plan" to activate plan mode. The agent will write out its steps.
- Review the plan. You can edit steps, delete steps, or add new ones.
- When the plan looks right, approve it. The agent will start working.
- To message a sub-agent directly, type @ in the composer and pick the sub-agent. Write your instruction and send it.
- Use the sandbox settings to control which files and networks the agent can access.
Tips, Limits, and Mistakes to Avoid
The biggest limit is still the agent's ability to run off and do the wrong thing. Plan mode helps, but you must use it. If you skip review, you risk coming back to something you didn't want. Set the review mode to "always" for important jobs.
When installing plugins, don't just click install. Check what each plugin asks to access. Curated shelves help, but you still need to be careful. It's the same habit you use with phone apps.
For developers using local models, remember that the local option is in the SDK for now. It's aimed at builders, but the idea of offline work is for everyone. If you need privacy or no internet, this is the way to go.
Finally, don't ignore the Windows sandbox. On a work laptop, admin rights are often blocked. The new sandbox works without them, so you can use the safe setup without waiting for IT.
Who Should Care About These Updates
These updates matter for anyone who uses AI agents in business. If you're a business owner, you can use plan mode to keep agents on track. You can build a team of sub-agents to handle research, writing, and checking. The plugin store makes it easier to connect antigravity to your existing tools.
For developers, the local AI models and SDK are huge. You can run agents offline, keep data private, and choose your own model. The Windows sandbox makes it safer on company computers.
Right now, most business owners haven't tried these features. That's a window. The people who learn early get to build their systems first. Julian's AI Profit Boardroom offers a 30-day roadmap, daily tutorials, and coaching calls to help you get started.
