Google Antigravity puts agents in the editor, the terminal and a real browser, then hands back the plan, the screenshots and the command logs so you can check the work before it lands. This is an independent community guide to what it does, where it falls short, and how to get it running.
walkthrough.md — plan, diff, the six screenshots it captured and the exact commands it ran. Read it before you approve anything.
One agent works across the editor, the terminal and a controlled browser.
Gemini 3 Pro, Claude Sonnet 4.5 and OpenAI's open-weight GPT-OSS.
Google describes preview rate limits as refreshing on a rolling five-hour cycle.
No card required while the public preview lasts. Preview terms can change.
Figures checked 24 August 2026 against Google's published preview documentation. Verify on the official site before relying on them.
Most AI editors stop at the file. Antigravity lets one agent write the code, run the command and then open the page to see whether the change worked — and keeps a record of each step.
A dashboard rather than a chat window. Each agent gets its own workspace, its own task and its own progress trail, so a long refactor is not competing with a quick bug fix for the same conversation.
The editor is built on the VS Code codebase, so your keybindings, themes and most extensions come across. The difference is that the agent can also run the command and read what came back.
// agent: added a retry budget, capped at 3
export const limiter = rateLimit({
windowMs: 60_000,
max: 240,
retryBudget: 3,
})
// terminal: pnpm test --filter billing
// 118 passing, 0 failing, 2.4s
This is the part other AI editors do not have. The agent starts your dev server, opens the page, clicks through the flow and takes screenshots. If the button does not work, it finds out before you do.
An agent that says "done" is asking for trust it has not earned. Antigravity attaches artifacts to every task so you can check the reasoning, the commands and the result separately.
Written before the work starts, in the terms of your codebase. Edit it, comment on a paragraph, or reject the approach outright — cheaper than reviewing the diff afterwards.
Screenshots of the running application at each step, tied to the action that produced them. Select any region of a screenshot and leave a comment; the agent picks it up as its next instruction.
Every command, in order, with the output it produced. The fastest way to find the moment a run went sideways.
Sub-tasks tick over as the agent works, so a long-running job is legible at a glance instead of being one opaque spinner.
Corrections you make are written back and retrieved on later runs, so the same house conventions do not need explaining twice.
Browser-in-the-loop work benefits most, because the agent can see the result. Work with no observable output gets less out of it.
UI changes have a visible result, which is exactly what browser control is for. The agent makes the change, reloads the page and shows you the state it reached.
Cross-cutting changes are where an up-front implementation plan earns its keep. Fix the approach on the plan, and the diff usually follows.
Teams juggling repositories get the most from the Agent Manager: separate agents, separate workspaces, one place to see what is waiting on a human.
Google published the thinking behind Antigravity. Whether the product lives up to it is your call — but knowing the intent makes the interface much easier to read.
Agent work is shown at task level rather than token level, with the smallest set of artifacts needed to judge it: the plan, the commands, the screenshots, the result.
Agents operate across every surface at once and keep going while you are elsewhere. The interface is built for asynchronous work rather than a turn-by-turn chat.
Comments attach to the artifact itself — a paragraph of the plan, a region of a screenshot — instead of being retyped as a new prompt.
Corrections are written to a knowledge base and retrieved on later tasks, so a convention you explained once does not need explaining again next week.
Short version below. The full feature table, with pricing checked on the vendors' own pages, is on the comparison page.
Multi-agent manager · browser control · artifacts as evidence
The only one of the three that can open your app and look at it. Best fit if you want agents doing whole tasks and proving the result. Still a preview, so expect rough edges and changing limits.
Agent, edit and ask modes · background agents
The most control over how the model edits your files, and the most mature at it. Strong single-developer choice if you prefer steering to delegating.
Cascade flow · previews and deploy · rules and memory
Built around keeping momentum in a large existing codebase, with deployment previews and a plugin ecosystem already in place.
It is free to use during Google's public preview, with rate limits rather than a bill. No card is required. Preview pricing is not a commitment — Google can change the terms, so check the official site before you plan around it.
Pick one bug with a visible symptom, let an agent take it end to end, and read the walkthrough it produces. Twenty minutes tells you more than any feature list.
The download link goes to Google's own site. We are an independent community publication, we do not host the software, and we receive nothing if you install it. See our disclaimer.