Getting Started with Google Antigravity: A Complete Guide to Agent-First AI Development

Google Antigravity is a free, experimental development platform that reimagines the traditional IDE around autonomous AI agents. Built on the open-source Visual Studio Code foundation, it goes beyond line-by-line autocomplete. Instead, it provides a 'Mission Control' where developers dispatch agents that can plan, write code, execute terminal commands, and even browse the web to complete engineering tasks. This shift changes the developer's role from writing and editing every snippet to acting as an architect or manager who orchestrates a workforce of digital agents. Developers can still drop into the familiar editor whenever they want to make direct changes.

Antigravity is available in preview, and you can sign in with a personal Gmail account. Download the installer for your operating system from the official downloads page and complete the installation. On first launch, the setup wizard lets you import settings from VS Code or Cursor or start fresh, choose a theme, and sign in to your Google account through the browser. For users on Linux subsystems or ChromeOS, community guides describe running the Linux build under WSL or ChromeOS Flex. Accept the terms of use to finish setup and you are ready to start collaborating with the agent.

During setup you configure how much autonomy the agent receives through two settings. The Terminal execution policy controls command execution: Off requires approval for everything outside an Allow list, Auto lets the agent decide when to ask, and Turbo auto-executes commands except those on a Deny list. The Review policy controls when the agent pauses for your sign-off on artifacts: Always Proceed, Agent Decides, or Request Review. These are bundled into four presets: Agent-driven development, Agent-assisted development (the recommended balanced option), Review-driven development, and Custom configuration. All of these settings can be changed at any time, even mid-conversation.

Antigravity splits its interface into two primary windows: the Editor and the Agent Manager. The Agent Manager is the default landing view and works like a dashboard for spawning, monitoring, and interacting with multiple agents running asynchronously. Key elements include the Inbox, which tracks all conversations, task status, and pending approvals; Start Conversation, which opens a new prompt; Workspaces, which are project folders you can add and switch between; the Playground, a scratch area for experimentation that can later be converted into a workspace; the Editor View, for inspecting and editing generated files directly; and the integrated Browser. Because agents run in parallel, a developer can assign several bugs or tasks at once rather than waiting on a linear chat.

When starting a conversation, you can choose the underlying model from the Model Selection dropdown. You can also choose between two modes. Planning mode lets the agent organize work into task groups, produce artifacts, and think through complex or research-heavy tasks thoroughly, producing more output. Fast mode executes tasks directly and suits simple, localized jobs such as renaming variables or running a few shell commands. Think of the choice as controlling the agent's thinking budget. Note that access to Gemini 3 Pro may be subject to usage quotas during preview, so you may see messages when free limits are reached.

When an agent needs to interact with the web, it delegates to a browser subagent running a model specialized for operating pages in the Antigravity-managed browser. This subagent can click, scroll, type, read console logs, capture DOM content, take screenshots or markdown snapshots, and record video. To enable it, start a conversation in the Playground with a prompt such as 'go to antigravity.google.' When prompted, click Setup, which opens the browser with an instruction to install the Antigravity Chrome extension. After installing the extension, switch back to the Agent Manager and grant the agent permission to proceed. Once authorized, the agent navigates the site on your behalf.

Artifacts address the trust gap that arises when an agent claims a task is finished. Rather than requiring you to read every line of code, Antigravity produces verifiable evidence. Typical artifacts include task lists and implementation plans that you can review, edit, and approve before work begins; code diffs showing exactly which lines change; screenshots capturing UI states before and after a change; browser recordings of dynamic interactions for functional verification; and structured test results. Together, these let developers confirm what was planned, what was done, and how it was validated, making agent output auditable and trustworthy.

To make the agent consistently follow your standards, Antigravity supports Rules, which are persistent instructions covering coding conventions and guidelines. Workflows are reusable, multi-step instruction sets that you can invoke with a single command, which is useful for repeatable tasks. Later updates added support for MCP (Model Context Protocol) Servers, letting agents connect to external tools and data sources, and Skills, which extend the agent's capabilities in Antigravity. Together these customizations help tailor the agent to your codebase and team practices so it produces output aligned with how you work.

Because agents execute shell commands to accomplish their work, controlling what they are allowed to run is essential. Antigravity lets you configure an Allow list of commands that may run without asking, and a Deny list of commands that are blocked outright. Combined with the terminal execution policy, this lets you permit routine, safe operations while requiring explicit approval for sensitive ones. Browser allow lists serve a similar purpose for web interactions. Establishing these boundaries before handing agents substantial autonomy is a practical safeguard for any real project.

Antigravity represents a shift from writing code line by line toward directing autonomous agents through a Mission Control interface, supported by artifacts that make their work verifiable. By completing setup, exploring use cases such as website generation and dynamic web application development, customizing behavior with Rules and Workflows, and securing commands with allow and deny lists, developers can begin integrating agentic development into their daily work. Because the platform is in active preview and evolving rapidly, refer to the official documentation, use-case gallery, and YouTube channel for the latest capabilities and best practices.