Google Antigravity Guide: The Agent-First IDE

Why this matters

Most AI IDEs are still VS Code with a chat panel bolted on. Google Antigravity is built around a different premise: the agent does the work, you direct traffic.

Announced alongside Gemini in late 2025, Antigravity has two surfaces — an Editor view that looks like a normal IDE, and a Manager view purpose-built for orchestrating multiple agents concurrently. You can have one agent refactoring your API layer while another builds a new UI component, both running asynchronously while you review work in the editor.

It's not a better autocomplete. It's a different category of tool — and it requires a different way of working. If you're evaluating whether to switch, Antigravity vs Cursor covers the concrete trade-offs.

The setup

Download the installer from antigravity.google — Windows, macOS, or Linux. On first launch you can import settings from VS Code or Cursor. Two things to configure before starting a real task:

1. Model — Antigravity defaults to Gemini 3.1 Pro. The picker also exposes Gemini 3 Flash (faster, cheaper) and Claude Sonnet. It's per-task, not global.
2. Index your codebase — Antigravity ingests your entire repo into a 1M token context window rather than chunked RAG. For large codebases, trigger this first. It takes a few minutes and makes everything afterward faster.

Step 1: Install and pick a model for your task

The model picker lives in the agent sidebar and the Manager view task launcher. Main options as of mid-2026:

• Gemini 3.1 Pro — default, best reasoning, right for anything architectural
• Gemini 3 Flash — 3–5x faster, lower cost, right for isolated changes
• Claude Sonnet — alternative for teams already in the Anthropic ecosystem

# Plan mode task prompt — Gemini 3.1 Pro
Model: Gemini 3.1 Pro
Mode: Plan

Refactor the authentication module to support OAuth 2.0 alongside
the existing email/password flow. Keep backward compatibility.
Write tests for both flows before touching any existing code.

Flash makes more assumptions and keeps moving. Pro stops and surfaces a plan before acting. Pick based on what you want the agent to do with ambiguity.

Step 2: Run an agent task end-to-end

Antigravity has two execution modes:

Plan mode — the agent generates a Plan Artifact first: a structured task list with the intended approach spelled out. You review it, edit if needed, then approve. The agent only acts after sign-off. Use this for architectural changes, cross-cutting work, or anything where a wrong assumption is expensive.

Fast mode — skips the planning artifact and executes immediately. Right for bounded tasks: fix this bug, update this component's styles, rename this variable. Both modes produce Artifacts as output — task checklists, implementation notes, browser screenshots — making review more structured than a raw diff.

# Fast mode prompt — scoped, low-risk
Update the Button component to use the design system's
`surface-raised` token instead of the hardcoded `#1a1a2e`.
Apply to all variants.

For anything beyond a single file, go Plan mode. The 90-second planning overhead is cheap compared to cleanup from a wrong assumption.

Step 3: Manage parallel agents in the Manager view

The Manager view is Antigravity's most distinct surface (Cmd+Shift+M). It shows all active agents across workspaces — each running in isolation, each with its own task and status.

Practical patterns:

• Background maintenance — dispatch a refactor to an agent, keep working in the editor, review when it pings complete
• Parallel feature work — one agent on the API endpoint, one on the frontend component — agents don't share state, so no collision
• Long-running investigation — point an agent at a bug with full repo context while you work on something else

Workspaces are isolated but share the codebase snapshot from dispatch time. Two agents editing overlapping files produce a merge conflict at review — the Manager surfaces this before you apply changes. Sequence tasks on overlapping files instead of parallelizing them.

Step 4: Use the browser tool for grounded research and UI verification

Antigravity's browser subagent can launch Chrome, navigate to a URL, interact with UI elements, and return screenshots as part of a task Artifact — built-in, no separate config.

UI verification — end your prompt with an instruction to open the local dev server and screenshot the result. The agent checks its own work visually, not just whether the code compiled.

Grounded research — include a docs URL in the prompt for any library the agent might not know well. The browser subagent fetches current documentation mid-task.

# Prompt using the browser tool
Implement a date range picker using react-day-picker v9.
Fetch current docs from https://react-day-picker.js.org before writing
code. After implementation, open http://localhost:3000 and screenshot.

This is different from Cursor's @Web mention, which injects a search result into the prompt context. Antigravity's browser tool runs mid-task execution — fetch, read, write, verify in one pass.

Common mistakes

Using Fast mode for complex tasks. If you're touching more than two files or more than one system, switch to Plan mode. Fast mode's lack of planning is a feature for small changes and a liability for large ones.

Skim-approving Plan Artifacts. The plan is where you catch the agent misunderstanding your intent before it acts. Approving without reading is the main source of "technically correct but completely wrong" agent outputs.

Parallel agents on overlapping files. Antigravity surfaces conflicts at review time, not before. If two workspaces will likely touch the same files, sequence them.

Over-specifying the procedure. These agents work best with goals, not step-by-step instructions. Describe the outcome and constraints, let the Plan Artifact show the approach. See prompting patterns for code for what good task prompts look like.

What's next

Antigravity is in public preview as of mid-2026 with all features free — pricing for general release hasn't been announced. Low-risk time to build the agentic workflow muscle before it's a paid tier.

For head-to-head tool comparison: Antigravity vs Cursor. For the underlying prompt craft that makes agent tasks land right: prompting patterns for code. New to the whole paradigm: what is vibecoding.