Antigravity’s Capabilities Tested: Two Experiments That Prove Google’s AI Revolution

‍

To evaluate the capabilities of Google’s latest tool, Antigravity, we designed two experiments to create software projects entirely from scratch. Here’s how the first experiment unfolded:

Experiment 1: Zerohost Firebase CMS

The goal was to create a CMS hosted on Firebase, capable of importing Webflow or WordPress websites, and recreate the entire website with a self-hosted CMS for streamlined content management. The reason is simple: Firebase is free, which is why it’s called Zerohost—no hosting fees required.

This project was already in progress. Our initial approach was straightforward:

• Break the project into smaller, object-oriented components.
• Use LLMs like Claude to generate each component.
• Assemble the pieces manually into a working system.

For this experiment on using Antigravity, we took a different approach. We fed the entire project requirement into Antigravity and let it generate the system from scratch—no manual breakdown, no piecing it together.

‍

The result

‍

‍

Later in this article, we’ll compare the outcomes of these two approaches to understand their strengths and limitations.

‍

Experiment 2: Reviving an Old Project with Antigravity

For the second experiment, I revisited an old project—a Bible journal web app—that I had originally attempted to build with Firebase Studio six months ago. Back then, the results were far from satisfactory. The process was clunky, and the outcome didn’t meet expectations (details here).

This time, I decided to rebuild the project using Antigravity, and the results were a pleasant surprise.

About the Project

The app is a Bible journal web app designed to offer a simple yet meaningful experience. Users can:

• Read Bible verses.
• Bookmark their favorite verses.
• Write journals tied to those bookmarked verses.

Sounds simple, right? But I decided to push it further.

Taking It to the Next Level with Antigravity

I asked Antigravity to do something unconventional:

• No server implementation. The app became a fully client-side experience.
• P2P sync. Journal and Bookmark data moves directly between browsers, bypassing servers entirely. This ensures total data sovereignty—users have full control over their data, without it ever touching a central server.

The result? A streamlined, innovative app that eliminates the need for traditional infrastructure while offering a modern, privacy-first experience.

‍

‍

What’s Good and What’s Not

What’s Good

1. Clean Software Architecture: The generated software architecture is well-structured, organized, and, most importantly, functional.
2. Accurate Integration with Google Cloud: The code for integrating and utilizing resources in Google Cloud and Firebase is precise and reliable. Unlike other LLMs, which often provide incorrect or incomplete guidance for Google Cloud integration, this one gets it right.

What’s Bad

1. Minor Bugs: While the bugs are small and relatively easy to fix, addressing them requires the expertise of a senior software engineer.
2. Basic UI: The user interface, though functional, lacks the polish and visual appeal provided by some other LLMs.

‍

Verdict: Antigravity vs. Other AI Vibe Coding Tools

Antigravity is undeniably powerful—it generates all the required features in one go. This is a significant win for Google. Instead of comparing it to other AI vibe coding tools, I found it more meaningful to compare it with my previous approach to using these tools.

The two approaches to "vibe coding" serve different purposes, each with its own strengths:

1. Top-Down AI Agent Approach (e.g., Antigravity):
   • Ideal for new, greenfield projects.
   • Excels in generating clean, well-structured software architecture.
   • Handles complex integrations, such as Google Cloud/Firebase, with precision and reliability.
   • Perfect for starting fresh with minimal setup effort and a focus on scalability.
2. Pre-Designed Object-Oriented Component Approach:
   • Best suited for large, pre-existing proprietary codebases.
   • Ensures strict backward compatibility, preserving stability in production workflows.
   • Critical for maintaining business confidentiality by reducing the risk of exposing sensitive code to LLMs.

Each approach has its place, depending on the project's needs. While Antigravity shines for building new systems from scratch, the component-based method remains indispensable for maintaining and enhancing legacy systems.

‍

Conclusion

AI agents are undeniably powerful, and with Google leading the charge, this comes as no surprise. As the creator of Google Chrome, and Google Search, Google has had access to vast amounts of cached web data across the internet (excluding the dark web) for years. Couple this with their foundational role in developing large language models (LLMs)—as introduced in the groundbreaking research paper Attention Is All You Need—it’s no wonder they’ve created a far superior AI agent tool compared to their competitors.

The real question is: why did they release it so late? Perhaps it’s part of a calculated business strategy—waiting for rivals to exhaust resources or burn through cash before striking decisively.

In my two experiments, I focused only on web apps and haven’t tested Antigravity’s potential for mobile app development yet. However, as we discussed in our earlier article, Bringing AI to Life: How Generative Agents Mimic Human Behavior with Memory and Reflection, AI is a transformative tool poised to reshape human society and the workplace forever.

The only remaining question is how soon. Antigravity offers a clear answer: the time is now.

‍

Google's Antigravity: https://antigravity.google/
‍

Remarks: Meanwhile, Experiment 1 will be fully developed and released as a community-supported open-source project.

‍

Experiment 2 result -- My Bible Journal

My Bible Journal: https://my-bible-journal.web.app/

‍
‍