Beyond Vibe Coding: Differentiated Workflows with Antigravity

Hello, my name is Alex, and I am a Developer Relations Engineer, usually based from Waterloo, Canada. And I'm Rodi. I'm also a Developer Relations Engineer, and I'm based out of San Francisco.

Let's start with the basics. What is antigravity and why should you care? There are many different ways to build with AI, but the harness, the tools, and the model can make a really big difference. With antigravity, the best of Google tools are coming together to create a unified harness and a powerful coding agent.

Instead of the coding agent just living inside of your IDE, your agent lives across multiple surfaces that the developer might spend time on. We include an agent manager. This is your simple hub, an agent-first view with a higher-level view than your code and your projects. It focuses on orchestrating multiple agents running across your code base. It has a single unified agent manager, and you can work across all your projects across your entire machine.

We also have an AI editor with all the bells and whistles that you're familiar with. It's a VS Code-based interface. It has a lightning-fast autocomplete, as well as a tab-complete model. It has an agent sidebar, where you can be able to have multimodal input and do so much more. And at any point, you can hit Command-E or the equivalent on Windows to get back to the agent manager and going back to the editor. There's also the agent-controlled Chrome browser, which lets you give the agent access to the richness of the web, and not only the context over specific things like the images and things being rendered, but you can also do complex things like have it complete tasks for you and be able to fill out forms, as well as to be able to run your automation tests.

Let's start. But before we get started, we need to install antigravity, which is very straightforward. You either download and follow the prompts for your system, or if you're like me, you just send that brew install antigravity command and grab a coffee. Then go through the installation onboarding wizard, sign in, and you are done. However, there is one step I like to spend some time on. The important thing for us to decide is the level of autonomy we'll give to the agent. You'll want to balance convenience versus control. Also, remember that better safe than sorry, so make sure you follow a few best practices. One, start defensively. Increase automation over time. And two, take your time to get comfortable with each model, each sub-agent, and each task. Three, keep sandbox enabled and only disable it if absolutely necessary.

Three, keep sandbox enabled and only disable it if absolutely necessary. And four, allow list actions and website mindfully. As long as you control those properties, it's probably safe. But if you're lurking on the deep side of the web, be mindful. These agents might be tricked. Like every other software engineering skill, AI steering takes some time before it feels intuitive. Enjoy the learning journey and do not rush it.

Anti-gravity offers two ways to build with the agent. See the difference for yourself. Presenting on the left side, fast mode. On the right side, the default planning mode. Fast mode is for building quickly and is especially useful with Gemini Flash model. Sometimes you just need to trigger a task to be completed, like building a quick mockup, a workflow, updating a doc, and you don't want to spend time going into too much detail.

As you can see, there is a lot more steering capability that you can find in the planning mode. This is great when starting to work on a more complex task. It can execute a lot of different searches, gather information for you, and allows you to fine-tune how the generation should flow. Both Flash and Pro models are decent choices for this mode. Using the planning mode, we can take the idea that we have and kick off a new project in the playground. The playground is a great place to validate ideas before committing to a full project.

The playground is a great place to validate ideas before committing to a full project. After we send the prompt, we will have a full overview of the app that gets generated with all the features that need to be built. And one by one, the agent will start to work on the tasks and make progress. At the end, it'll generate a walkthrough with steps on how to run it and verify that it's working. At each stage, the agent generates artifacts for us that ensure that there is visibility into what is being created. And for things like the overview, it can offer the human in the loop an ability to comment and provide feedback at every step in the way to steer the output.

We also offer image generation. This is for building UI designs as well as assets in the application. It also has a very good image understanding, like being able to understand screenshots and browser recordings as well as many other types of error messages and things that you can provide to the agent. This is the beautiful part about AntiGravity, where it's bringing all of Google's products together with a great synergy. And Nano Banana Pro has been available since day one in AntiGravity, so you can rest assured that new model capabilities will come to AntiGravity so that you can take advantage of them. Gemini 3 has greatly improved on the design output and code generation. And in AntiGravity, you can take advantage of it with all the ways that you might interface with a designer or asset generator. You can directly attach images into the chat and with detailed instructions on how to implement it. You can also use MCP servers like Stitch or Figma to directly grab design assets and implement them as code.

AntiGravity recently introduced support for skills. They are a way to add deep and narrow context on specific tasks that require knowledge that may not be in the training data or easily found with search grounding. When you ask the agent to do something, it can search the skills that it has for possible matches and then load them into context and provide a much better output. Skills can also include scripts, resources, examples or even instructions on how to call a specific MCP server. Talking about scripts, what happens if a script that the agent runs ends up blocking and waiting for the input? Not much really. AntiGravity will wait for a few seconds and then it will take action. Be it sending a key combination, typing text, interactive commands, anything. Even better, due to its asynchronous nature, the terminal agent can run concurrently, starting multiple servers and scripts simultaneously, making complex project development much easier to automate. And my favorite is the browser agent. It's like that universal multi-purpose tool that I want to keep around with me just in case. Might not be the fastest, but hey, it's the most universal thing ever. Also, it's an amazing time saver trick, especially when dealing with legacy code. Each time when a UI frame gets implemented, you can let AntiGravity go ahead and run and poke around, finding and fixing those UI bugs before I even notice them. What's even better, it will record its interaction and the most important steps presented in that final report. Remember the agent manager? Essentially, I get something I can quickly check and present to my entire team, and this can also be a great start for the next sprint or feature.

We have all this automation, but I sometimes want a manual gearbox that gives me full control of how I'm driving, not just steering. In these moments of inspiration, I take control and flow using the keyboard. This collection of delightful top features makes me smile on those evenings when I want to craft some code, just like in the old good days.

I think we're almost there. When you're ready to move beyond an agent-led chat in your editor, you then can start to work with the agent manager that is included inside of AntiGravity. This gives you the visibility into all the agents on your machine and your projects, and we're hoping that it will help change the way that you work. Anything that requires input from you as the human will surface here, and you can make decisions and actions on how to proceed.

There is so much more that we haven't talked about, and we hope that this just gets you curious to try out a different type of agent-led experiences. There's a lot of updates and features rolling out constantly, and we're monitoring your feedback, always looking for ways to improve. Make sure to follow us on social for ways that we're posting every day and techniques that you can actually apply to your workflow. And make sure to head over to AntiGravity.Google to get started. Thanks.