Google Antigravity puts you in charge of a crew of coding agents

Somewhere around the third time you pasted a task into a chat window, waited, copied the answer back into your editor, and then noticed the answer quietly broke two other things, a thought probably crossed your mind: why am I the messenger here? That frustration is the starting point for Google Antigravity, and it explains the strange direction the product took in 2026.

Google launched Antigravity in November 2025 as its agentic development platform, then relaunched the whole thing as Antigravity 2.0 at Google I/O in May 2026. The headline change: version 2.0 is a standalone desktop app for macOS, Linux, and Windows, with no IDE required. You can visit the official Google Antigravity page to see the current shape of the product, because the team ships updates fast enough that any single screenshot goes stale within weeks.

That sounds like a small packaging change until you think about what it implies. Most AI coding tools position the assistant as a helper inside your editor: you type, it suggests, you accept. Antigravity flips the seating arrangement. The app treats agents as the workers and you as the person who assigns work, checks results, and decides what gets merged.

Here is the part that made me sit up: the main agent can spawn subagents. When you hand it a task with several independent pieces, it registers small helper agents, gives each one a scoped piece of the job, and runs them in parallel with isolated workspaces. The main agent keeps its own context clean instead of turning into a mud puddle of half-read files. Google's I/O 2026 write-up describes this with native Git worktree support, so each subagent can poke around in its own background folder without stepping on the others. When the work finishes, the worktrees get cleaned up. If you have ever watched a single long agent session degrade into confusion, this design will read like a reasonable response to that exact failure mode.

Another feature deserves special attention because it quietly changes what "using a tool" means. Scheduled Tasks let you set up cron-style recurring agent jobs. The official docs give examples like a daily pull request digest, a deployment check, or a monthly architecture report. You configure it once, and the agent runs on schedule without anyone typing a prompt. That moves the product from "assistant you visit" toward "coworker with a calendar." Whether that is comfortable or creepy probably depends on how much you trust your own task descriptions, which is a fair self-test before letting software loose on a schedule.

The way work gets organized is also a departure from the usual. Instead of grouping conversations by repository or workspace, Antigravity groups them into Projects, and a Project can span multiple folders with its own scoped settings and permissions. There are also standalone one-off conversations when you just want a quick answer without opening a whole project. Security controls sit around all of this: agents ask for approval on sensitive actions and stay inside bounded folder access unless you say otherwise.

What you actually get in the box

The Antigravity family is bigger than the desktop app, and it helps to know the pieces before deciding where it fits in your setup:

• Antigravity 2.0, the standalone desktop command center for supervising agents.
• Antigravity CLI, which replaced the Gemini CLI for consumer terminal use in mid-2026.
• An SDK for building custom agents on the platform.
• IDE extensions, so agents can also show up inside VS Code, Visual Studio, JetBrains IDEs, and Zed, several of them in preview.
• Remote Control, added in August 2026, which lets you drive a desktop session from a browser.

That last one is worth a second look if you bounce between machines. Remote Control means you can start an agent job on your desktop and keep an eye on it from a laptop or tablet browser. For anyone who has ever left a long build running and walked away muttering "text me when it's done," this is Google's version of texting you.

Under the hood, the platform pulls from a menu of models rather than betting on one. The pricing page lists Gemini 3.8 Flash, 3.7, and 3.6 Flash, Gemini 3.1 Pro, Claude Sonnet and Opus 4.6, and the open-weight gpt-oss-120b. Google added Gemini 3.8 Flash availability on September 1, 2026, which tells you the model lineup moves often. For a reader, the practical takeaway is that you are not locked into one vendor's taste in models, and Google clearly intends to keep refreshing the roster.

What it costs, in plain terms

The pricing shape is simple at the entry point. The Individual tier costs $0 and includes agent model access, unlimited tab completions, unlimited command requests, and what Google calls basic weekly rate limits. That word "basic" is doing honest work there. Paid capacity rides on Google AI Pro and Google AI Ultra subscriptions, which buy more generous rate limits and a flexible pool of AI credits. Google does not publish a separate Antigravity-only price. For organizations, access goes through Google Cloud, with consumption-based pricing and availability inside select Gemini Enterprise app subscriptions, plus admin and spend controls in the Gemini Enterprise admin console.

If you are the kind of person who burns through free tiers in a weekend, set expectations accordingly. Widespread user reports describe heavy agentic sessions eating quota quickly on the free tier, which tracks with how much compute a swarm of parallel agents plausibly consumes. The free tier is genuinely useful for trying the workflow. Sustained daily agent farming is what the paid plans are for.

Where it runs, and what that means for your data

Antigravity is a cloud product. There is no local model runtime, so if your requirement is "nothing leaves my machine," this is not your tool. Models run in Google's cloud, and agent sessions coordinate through Google's infrastructure. Privacy posture differs between consumer accounts and Workspace or Cloud access according to Google's terms, and those terms are exactly the kind of document that changes, so check the current terms rather than trusting a blog summary, including this one. On the plus side, the security model around agent actions is more careful than most: interactive approvals and bounded folder access are defaults, not opt-ins.

One honest gap for some readers: no local mode also means no offline work and no freedom from rate limits. If you want a local-first agent platform, alternatives like Ollama paired with an open-source agent framework live in a different corner of the map.

Who should try it, and who should wait

This product fits a particular kind of person well: someone with a real backlog of small-to-medium coding tasks, some tolerance for supervising software instead of typing every line, and a willingness to review artifacts before accepting them. Solo developers juggling multiple side projects are a natural fit, because Projects span folders and keep contexts separate. Teams that already live in Google Cloud get the smoothest enterprise path. And if your work has a recurring rhythm, weekly reports, standing checks, regular digests, Scheduled Tasks alone might justify the install.

Who should wait? Anyone whose entire workflow fits inside one editor window and who enjoys writing every line themselves; the command center model buys you nothing there. Anyone on a strict "cloud code touches nothing" policy should also hold off until they have read the terms carefully. And if you already pay for a competing agentic tool that works, switching costs are real.

How it compares with the tools you may already know

Cursor keeps the editor at the center and adds increasingly capable agents to it, which makes it a gentler transition for people who love their current setup. Claude Code works from the terminal and appeals to developers who want an agent that lives where their scripts live. Devin positions itself as a hands-off teammate for whole tickets. Replit Agent bundles hosting and deployment into the loop. Antigravity's distinguishing bet is the supervision surface: it wants to be the room where several agents work at once, not a better autocomplete or a single clever terminal session. If you have more tasks than attention, that framing is worth a test drive. If you have one task and full attention, any of the others will feel lighter.

There is also a practical difference in how work gets handed to you. Antigravity surfaces artifacts, plans, screenshots, and task results that you review in the app, so the review step is built into the product instead of being something you improvise in a diff view. Google's own docs frame this as the intended loop: agents produce verifiable work products, and you approve what you have actually inspected. Skim the artifact, click approve, and the whole model collapses into expensive automation theater. Read them properly and it genuinely multiplies what one person can ship in a week.

Getting started without getting burned

A sane first week looks something like this. Install the desktop app, sign in with a Google account, and start with one small, well-defined task rather than a grand architecture migration. Give the agent a bounded folder. Watch how it plans, what artifacts it produces, and how often it asks for approval. Try one Scheduled Task for something low-risk, like a daily summary of your open pull requests. Only after a few days of that should you consider parallel subagents on real work. The people who bounce off agentic tools usually skip the supervised warm-up phase and then blame the tool for the chaos that follows, which is a bit like blaming a bicycle for a mountain.

Worth knowing before you commit: the platform changed packaging once already, from IDE-centric to standalone app, and its CLI replaced the Gemini CLI in 2026. Google is moving fast here, which is exciting and slightly exhausting. Feature names and menu locations in any Antigravity article, including this one, have a shelf life. The official feature documentation is the place to check current behavior before you build a workflow around a specific capability.

The short version: Google Antigravity is a real, actively developed bet that the future of coding work looks like a well-run control room rather than a very fast typist. It costs nothing to try, it runs on the machines you already own, and it handles the multi-agent, scheduled, supervised style of work that single-window tools struggle with. Try it when your to-do list is longer than your day. Skip it when the opposite is true, and spend the time writing code instead of supervising a small digital staff. And whenever you are between coding sessions, you can always download M&M POS if you also run a shop floor that needs a point of sale that simply behaves.