Google just shipped Anti-gravity 2.0, and it's not another chat interface bolted onto a code editor. It's a standalone desktop app built on the premise that AI agents should be running the project while you direct traffic.
That's a different model from what most of us are used to. Right now, the typical workflow is: you write code, you ask an AI to review or extend it, you review what it suggests, you repeat. Anti-gravity 2.0 flips that. You orchestrate multiple sub-agents, each handling a different part of the project simultaneously. Backend code in one agent, database work in another, UI in a third. They run in parallel.
Voice commands and work trees
You can issue instructions by voice. That sounds minor until you're deep in context-switching and you just want to say "add error handling to the auth flow" without stopping to switch windows and type. It's a small thing, but it adds up.
The more interesting feature is separate work trees per task. Each agent gets its own environment to test in, so they're not stepping on each other. The main friction with current AI tools is that everything is sequential. You wait, you review, you prompt again. Work trees break that bottleneck. Multiple agents can push forward at once, and you review when they're done rather than babysitting each step.
Image generation in the same workflow
For front-end work, the built-in image generation is practical. You're not jumping to a separate tool to generate a placeholder or mock up a UI element. Whether you actually need this depends on what you build, but having it in the same environment removes a context switch that happens more often than it should.
The background scheduler changes something
This is the part I keep coming back to. Anti-gravity 2.0 has a scheduled task system. You write a recurring prompt, set a schedule, and the agents execute it in the background even when the app is closed.
That's a different mental model entirely. It's not AI as a smart autocomplete. It's AI as a process running on your machine while you're not watching. Agents working overnight on tasks you configured before logging off is either a genuinely powerful workflow tool or something you'll want to audit carefully, depending on what you're having them do. Probably both, honestly.
The scheduler alone puts this in a different category from Copilot, Cursor, or anything else in the current crop of AI coding tools.
Gemini 3.5 Flash under the hood
The tool runs on Google's Gemini 3.5 Flash model. The demo walks through building, testing, and deploying a functional web application with minimal manual input. Fast completions, parallel agents, and something working at the end.
Who actually reaches for this
Anti-gravity 2.0 is not trying to be a better editor. It's trying to replace the coordination overhead. Solo developers juggling multiple concerns on a project are the obvious fit. You can delegate the parallel tracks and just handle reviews and decision points.
On a team with clear ownership, the picture is murkier. You'd need to think about where it fits in an existing workflow rather than assuming it slots in cleanly.
The background scheduler is not a feature you integrate and forget. It changes how you think about what AI is doing on your behalf, which is a conversation worth having before you fully hand it the keys.
