As noted in previous posts, the web chat interface can be a cumbersome method of interacting with an LLM for the purpose of web development tasks such as adding a CSS file to an HTML page to make it render nicely on smartphones.
What if we try artificial intelligence as a companion to a programmer’s integrated development environment (IDE)? Google’s $2.4 billion baby, Antigravity, is the contender for this post, a collaboration between John Morgan and Philip Greenspun.
Installing Antigravity immediately presents the new user with a challenging question: Review-driven development vs. Agent-driven development. There is no explanation on the setup screen as to what the difference between these two might be:
Antigravity’s default user interface presents the developer with three panes: a file explorer on the left, a code editor in the middle, and an LLM chat interface on the right. We start this experiment by opening a folder we’ve created named bh-antigravity-gemini that contains the HTML for the decidedly archaic Berkshire Hathaway homepage and proceeding to write a prompt asking to generate CSS and edit the page’s HTML to make it responsive:
Antigravity offers a variety of models to choose from but for the purposes of this experiment we’ll use the default which is Gemini 3 Pro (High) in Planning mode. Planning mode causes Gemini to come up with plans before implementing them. After submitting the prompt Gemini begins to think through a plan:
As Gemini works to develop a plan Antigravity presents a Task document showing the steps to be taken:
Once Gemini has come up with its plan Antigravity presents us with an Implementation Plan document outlining the proposed changes as well as a verification plan:
Since the plan sounds sensible we instruct Gemini to implement it:
Once Gemini completes the task Antigravity presents us with a Walkthrough document with an overview of the results of completing the task:
Below the message announcing the completion of the task Antigravity presents us with a list of files that have been changed including counts for lines added and removed. Just below the list is a button to “Accept all” changes. We want to inspect the changes for ourselves first so we select the berkshire-hathaway.htm file to see what was added. Sure enough, Gemini has kept the changes to a minimum as requested and simply inserted a viewport <meta> tag and a <link> to incorporate the newly created stylesheet:
We accept the changes in the HTML and move on to examine the newly created style.css file. The CSS includes comments indicating that Gemini has tailored it to match the styling present in the original HTML while implementing the proposed changes to make the page responsive:
Satisfied with its suggestions we accept the changes and proceed to view the results. The images below are the original page as viewed on desktop followed by the new page as viewed on desktop. As promised Gemini has kept the desktop styling largely the same with the primary difference being greater spacing between links:
The images below are the original page as viewed on mobile followed by the new page as viewed on mobile. The new CSS has transformed the page such that it is no longer rendered as it would be on desktop but instead presents the links as a single column with text that is readable and easy to select without needing to zoom.
The quality of Gemini’s code generation in this experiment appears comparable to what we saw from it in our previous comparison of ChatGPT, Claude, Gemini, and Grok. This is not surprising as Antigravity is merely another interface for a developer to interact with an LLM such as Gemini.
Our conclusion is that having AI integrated tightly into the IDE where the chat log is presented and persisted alongside the code and the model can be quickly directed to analyze files within the codebase without the need to paste them into a chat window or manually upload them via a web interface saves time and leads to a more organized and less fragmented experience. We also find that the IDE’s ability to highlight proposed changes and allow us to accept or reject them and edit our files in place achieves a level of integration into the development workflow that a web chat interface cannot match.