Antigravity Playground and Multi-Agent Collaborative Development

Introduction

Antigravity offers a collaborative Multi-Agent working mode and a Sandbox (Playground) mode for demonstration and development purposes.

Use the shortcut Ctrl + E to open Agent Management.

Playground

Antigravity provides a Sandbox (Playground) mode that can be used for more radical testing or for joint testing of multiple solutions, similar to what I will demonstrate below. I will create a traditional login page and a glass-morphism dynamic login page, where the dynamic login page will be rehearsed in the sandbox.

All of its code files are not in the current project, but rather in the playground folder mentioned in the previous article.

In this way, users can be provided with one or more additional demonstrations for continuous testing. To use it, simply click on “Add Conversation” in the Playground.

Agent Collaboration Mode

Since the design of Antigravity is to allow the user to become a project manager who can assign tasks to Agents (treating them as employees), you can set up multiple tasks to run in parallel in the Agent Manager.

Case Demonstration

Introduction

Using a simple login page as a reference, I pre-set Skills and Workflows in the project to organize the basic code style and coding steps respectively.

Skill

To demonstrate the difference in this project, I used Gemini to help me set up two Skills:

Frontend Skill: Basic Style: Focuses on generating production-grade, responsive, and A11y-compliant frontend style code.

Backend Skill: Fast API, add tag-hsu before log: Focuses on generating robust, secure, and high-performance backend logic and API interfaces.

Workflow

A linear workflow. Step 1 dispatches to the frontend Agent to generate UI and environment configuration; Step 2 uses the frontend code as context and dispatches to the backend Agent to generate supporting services.

Steps

1. Switch the project to your own project

   

2. Create a main task interface responsible for assigning work tasks. I constrained this by referencing the workflow mentioned above. As seen in the image, it actively reads the skill files for design.

3. After the main task creation is complete, different folders are designed within it. Next, you need to set up the Frontend and Backend Agents separately. I directly told them their identities and let them come to this main task list to organize their work.

Frontend Prompt:

Backend Prompt:

4. The basic setup ends here. You can switch to the inbox to monitor the status of all tasks, including our sandbox status. Subsequent work involves gradually checking their running status and permissions for various commands. If you feel there are no problems for now, you can also continue to develop a third style in the sandbox.

5. Since the project is relatively simple, the two formats, virtual environments, and backend startup will be completed very quickly. The left side is my default style, and the right side is the style constrained in my sandbox. You can see that it starts referencing from the installation location.

6. Project verification: Starting the backend and frontend, you can separately see that the frontend content and backend output meet the requirements.