Learn the basics of Google Antigravity

The following step-by-step instructional article is based on the demonstration of using Google Antigravity for agent-assisted development, as shown in the provided video.

How to Use Google Antigravity for Agent-Assisted Development

Google Antigravity, a product by Google DeepMind, combines an editor, an agent manager, an orchestration system, and a browser into a single product surface [00:05]. It is designed to help developers build applications efficiently using an AI agent.

Step 1: Get Started and Configure

  1. Open the App and Sign In: Launch the Anti-gravity application and proceed through the onboarding [00:20]. Choose your preferred display mode (e.g., Default Light Mode) [00:25] and sign in using your normal Google account [00:32].

  2. Select Development Mode: The video recommends choosing Agent Assisted Development [01:46]. In this mode, the Large Language Model (LLM) agent automatically determines whether to execute a simple task or pause to ask for user input on more complicated matters [01:52].

Step 2: Understand the Three Core Surfaces

Antigravity operates across three main integrated surfaces:

  • Agent Manager: This surface is used to manage and create agents across all of your workspaces [00:50].

  • Anti-gravity Editor: This is the traditional code editor environment. You can enter the editor at any time by pressing Command+E or the "open in editor" button. It features tools like tab autocomplete and an agent sidebar [01:05].

  • Browser: This is a dedicated offering that incorporates the agent directly, allowing it to test your features by clicking, scrolling, and interacting with the screen [01:19].

Step 3: Create a Workspace and Define the Task

  1. Add a Workspace: In the left sidebar, add a new workspace. For a local project, select Open folder and choose or create a project directory (e.g., flight tracker) [02:18], [02:28].

  2. Assign the Project: Provide a detailed prompt to the agent, outlining the application's required functionality, technologies, and specifications. For example, you might ask it to "Build me a flight lookup next.js JS web app where the user can put in a flight number..." [02:38].

Step 4: Monitor and Review Agent Artifacts

As the agent begins working, it generates artifacts—markdown files used to track its progress, perform research, and generate findings [03:30]. Check the right sidebar to monitor these:

  • Task List: A list the agent uses to keep track of its own progress [03:54].

  • Implementation Plan: This is a key document that details the agent's research and what specific changes it intends to make to your codebase before it starts implementation. You should review and approve this plan before changes are applied [04:04], [04:37].

  • Walkthrough: This final report is generated at the end of a task. It communicates what the agent did and provides verification steps, which may include screenshots, terminal commands, or unit tests, to prove the task was completed correctly [04:21].

Step 5: Leverage Parallel Tasks and Contextual Comments

  1. Start Parallel Work: You can return to the inbox/home screen to spin up new, parallel conversations and delegate background tasks to the agent [06:01], [06:10]. Examples include:

    • Research: Asking the agent to look up an API's documentation and use an API key to execute a curl request, providing sample data and a detailed research artifact [06:21], [08:21].

    • Design: Using integrated models (like Nano Banana) to generate logo mockups for the application [06:43].

  2. Provide Contextual Instructions: When reviewing an implementation plan or an agent's progress, you can highlight text and leave comments, similar to a Google Doc. The agent is smart enough to take these pending comments into consideration and inject them into its workflow at an appropriate stopping point [07:39], [08:56].

Step 6: Integrate and Finalize the Code

  1. Manual Code Integration: Once the agent has completed an auxiliary task (like creating an API utility file), you can move to the Editor [09:43]. The editor is context-aware and will use its tab autocomplete feature to suggest replacing mock data with the newly available live data utility [10:22].

  2. Request Final Features and Verification: Assign the agent the final feature (e.g., integrating the app with Google Calendar) and request that it test this with the browser and show me a walkthrough [11:34].

  3. Watch Real-Time Testing: The agent will then open the Browser (indicated by a blue border) and interact with the UI, clicking and testing the new feature to ensure it works, before generating the final Walkthrough report [12:23], [12:40].

  4. Commit the Code: When all tasks are complete, use the Generate a commit message button, which uses the conversation history and files to create a contextual commit message, and finalize the code [13:13], [13:24].

Google Antigravity, APIFrancesca Tabor