How to Use Google Antigravity for AI Coding

TL;DR

Google Antigravity supports a repeatable coding workflow: plan the work, let AI agents build it, test the application through a browser agent, review the resulting report, and iterate. Its Agent Manager can coordinate several conversations and workspaces simultaneously, while model selection lets users process requests with Gemini, Claude, or GPT models instead of relying on a single model family.

Transcript

Google's anti-gravity just released and in this video I'm going to show you how to use it, how to set it up on your local machine and I'm also going to show you three major features that you need to know if you're using anti-gravity here to do vibe coding. So first thing first I'm going to show you the one of the feature which is the agent manager.... Read More

Key Insights

• The Agent Manager is an orchestration interface for running multiple AI agents and conversations across different workspaces, allowing codebase analysis, design research, application testing, and implementation work to proceed at the same time instead of forcing every request into one sequential conversation.
• Google Antigravity setup includes downloading the application for the operating system, completing onboarding, optionally importing VS Code settings, selecting an editor theme, choosing the recommended agent-assisted development option, and accepting the default configuration used in the demonstration.
• Conversation context can include referenced project files or folders, attached files, and images, giving an agent specific source material for its assignment. The editor also provides different conversation modes, including a planning mode and a faster mode intended for direct execution.
• The browser agent works by opening an application, navigating its pages, taking screenshots, and analyzing the visible interface. It can examine both public landing pages and authenticated application areas when it receives the necessary navigation instructions and login credentials.
• Browser-generated findings can be organized into a detailed review report containing screenshots for individual pages, identified problems, and UX recommendations. Comments can then be added to the report so the agent can revise its analysis and continue the improvement process.
• Multiple model support allows a user to choose among Gemini, Claude, and GPT models for a new request. The demonstrated workflow therefore does not require every coding, research, planning, or interface-analysis task to be handled by a Gemini model.
• The recommended development cycle is to provide a build prompt, use planning mode to create a task list, give those instructions to a large language model, test the resulting application with the browser agent, and repeat the cycle based on the generated report.
• Parallel conversations can divide work by objective within the same workspace or across separate projects. The demonstration assigns agents to understand a codebase, assess landing-page conversion opportunities, research interface themes, inspect bookkeeping pages, test transaction matching, and implement a new project section.

Questions & Answers

Q: How do you set up Google Antigravity locally?

Download Google Antigravity for the operating system and open the application to begin onboarding. In the demonstrated Mac OS setup, the user imports VS Code settings, keeps the Solarized Light theme, selects the recommended agent-assisted development option, and accepts the default configuration. Completing those steps opens the Antigravity coding editor, where the Agent Manager, conversation modes, model selector, and browser capabilities become available.

Q: What is the Agent Manager in Google Antigravity?

The Agent Manager is an orchestrator for multiple AI agents, conversations, and workspaces. It can keep different projects visible, open new conversations, and allow separate assignments to run at the same time. In the demonstration, agents examine a codebase, evaluate a landing page, research interface themes, and test a bookkeeping application, reducing the need to wait for one task before starting another.

Q: How do you open and use the Agent Manager?

The Agent Manager can be opened with Command E or by selecting the visible Open Agent Manager control. From there, users can open an existing workspace or add another one, start a conversation, reference files or folders from the project, attach files or images, choose a conversation mode, select a supported model, and submit a focused request for the agent to process.

Q: How does the browser agent test a web application?

The browser agent opens the application, navigates through relevant pages, captures screenshots, and analyzes the interface or behavior it encounters. The demonstration asks it to inspect a landing page and an authenticated bookkeeping application, including receipts, transactions, statements, and matching features. It then records problems, evaluates the user experience, and produces recommendations supported by screenshots from the reviewed pages.

Q: How can Google Antigravity evaluate a landing page?

An agent can receive the landing page file as context and be asked to identify opportunities for improving conversion. It can also open the page in a browser, capture screenshots, research suitable UI themes, compare candidates, and make design decisions. In the demonstration, this process produces an updated page with recognizable customer names, calls to action, an education footer, and selectable light styling.

Q: Which AI models can be used in Google Antigravity?

The demonstrated model selector includes Gemini 3 Pro as well as Claude and GPT options. A user can choose a model when starting or configuring a conversation, so requests are not restricted to Gemini models. The transcript shows this selection being used for tasks such as codebase analysis and adding a project section to a static landing page, without claiming that one model is always preferable.

Q: What workflow improves accuracy when coding with Google Antigravity?

The presented workflow begins with a clear prompt describing what should be built. Planning mode then converts the request into a task list, which is passed to a large language model for implementation. After the application is built, the browser agent tests and analyzes it, creates a report describing results and problems, and supplies feedback for another cycle of revision until the application meets the intended requirements.

Q: How can Google Antigravity review an authenticated application?

The agent can be instructed to open the application and log in with provided credentials. It can then capture the dashboard and subpages, navigate through areas such as receipts, statements, transactions, and matching, and try relevant interactions. The requested output can include failures, page-specific findings, screenshots, and UX recommendations, assembled into a review document that supports comments and subsequent revisions.

Summary & Key Takeaways

• Google Antigravity is installed locally and configured through an onboarding process that can import VS Code settings, preserve a preferred editor theme, and enable agent-assisted development. The demonstrated setup uses the recommended agent option and default configuration before opening the coding editor and accessing its tools for planning, execution, browser testing, and review.

• The Agent Manager serves as an orchestrator for multiple AI agents, conversations, and project workspaces. Users can reference project files or folders, attach supporting material, choose planning or direct-execution modes, and assign separate tasks concurrently. Demonstrated tasks include codebase analysis, landing-page conversion review, design research, and authenticated bookkeeping application testing.

• The browser agent can open applications, capture screenshots, navigate authenticated pages, test interactions, and document usability problems. Its findings can become a detailed report containing page-specific screenshots, issues, and recommendations. Combined with selectable Gemini, Claude, and GPT models, this supports an iterative workflow of planning, building, verifying, reporting, and revising changes.