Google Antigravity 2.0 is a desktop platform for working with AI agents on software projects. Google describes it as a central place for launching, monitoring, and coordinating agents, including agents that can work on tasks in parallel. This guide presents a safe first workflow for using Antigravity with a small project.
Install Antigravity 2.0
Begin with the official Google Antigravity codelab. It describes Antigravity 2.0 as a standalone application for macOS, Linux, and Windows. The wider ecosystem also includes an agent-focused IDE, a command-line tool, an SDK, and extensions for other editors. This guide focuses on the standalone application, which Google says can manage multiple local agents in parallel and run scheduled tasks. (Google for Developers)
Download the application for your operating system, run the installer, and sign in with Google. The codelab says the personal setup uses a personal Gmail account. Business users can instead connect an enterprise Google Cloud project and related licenses or billing. (Google for Developers)
Create a test project
After signing in, create a small practice project. A simple web page, a Python utility, or a short JavaScript application is enough. Starting small lets you learn how the agent behaves before giving it a large codebase or access to production data.
Create a new folder and add a short text file that explains the project. Include the application's purpose, the language it uses, how it should be tested, and any files the agent must not change. These notes give the agent useful context and make its decisions easier to review.
In Antigravity, create a project and select the folder or folders it should use. Google's codelab says a project defines the environment and scope available to an agent, including its files, tools, and permissions. It also says projects have isolated agent settings, so separate projects can use different security configurations. (Google for Developers)
Once the project is open, start a new conversation and keep the first request narrow. For example:
Inspect this project and describe its files, purpose, and likely entry point. Do not edit anything yet. List any missing information before proposing a change.
This first request separates inspection from modification. Read the response and check whether the agent correctly understands the project before asking it to edit anything.
Delegate one coding task
After reviewing the agent's inspection, give it one small task with a clear success condition. For example:
Add input validation to the contact form. Keep the existing layout, reject empty names and invalid email addresses, and add tests for valid and invalid input. Explain each file you change.
Good requests define the result without prescribing every line of code. State what must remain unchanged, identify the expected behavior, and ask the agent to explain its work. Avoid asking it to “improve everything” in one session because broad requests are harder to check.
Google describes Antigravity as an agent-first development platform intended to help developers move from an idea to an application. Its desktop application supports agent interaction, parallel tasks, dynamic subagents, scheduled tasks, and integrations with Google AI Studio, Android, and Firebase. (Google)
Review changes and permissions
Before accepting a change, inspect the files that the agent modified. Compare the result with the original requirement, look for unrelated edits, and read any generated tests. If the agent reports a failed test, ask it to explain the failure before requesting another change.
Keep permissions narrow during early experiments. Do not give a practice agent access to cloud credentials, private customer data, payment systems, or production deployment tools. If the application asks for permission to run a command, access a service, or change files outside the project, pause and confirm that the action is necessary.
For browser or interface testing, Google's official codelab demonstrates an Antigravity CLI workflow that uses agent skills, BrowserMCP, and Playwright with a React demonstration application. The codelab shows how an agent can receive a plain-language testing request and interact with a browser, but the example is intended for a local test application. (Google for Developers)
Test before expanding scope
Run the project's existing tests after every meaningful change. If the project has no tests, ask Antigravity to propose a small test plan before it writes new code. Review the plan, then request only the tests that match the behavior you need.
When the first task works, repeat the workflow with a slightly larger change: inspect, plan, edit, test, and review. Keep a short record of the prompts that produced useful results. These notes can become project guidance for future sessions and help your team identify which tasks are safe to delegate.
Antigravity 2.0 is designed for parallel agent work, but parallel execution increases the number of changes that must be reviewed. Begin with one agent and one project. Add more agents only after you understand how the platform reports conflicts, permissions, test failures, and unfinished work.
References
- Google. (2026, May 19). Building the agentic future: Developer highlights from I/O 2026. https://blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-developer-highlights/
- Google for Developers. (n.d.). Automated UI testing with Antigravity (AGY) CLI, BrowserMCP, Playwright, and the Browser Agent. Retrieved November 19, 2026, from https://codelabs.developers.google.com/agentic-ui-automation-with-antigravity
- Google for Developers. (n.d.). Getting started with Google Antigravity. Retrieved November 19, 2026, from https://codelabs.developers.google.com/getting-started-google-antigravity
