Google Antigravity: building software with agents

What it is

Google Antigravity is Google's agent-based development environment, built around AI agents: AI that works through a task in steps, using tools, rather than answering once. You describe a goal and an agent plans the work, edits the code, runs commands in a terminal (the text window where programs are started) and tests the result in a browser.

It comes as a desktop App for overseeing several agents, a CLI (command line tool), extensions for common code editors plus a standalone editor, and an SDK (a toolkit for building your own agents). It is also built into the Gemini app for Mac. This is as of November 2026; check the maker's page.

Two ways of working matter most. In the editor, you work alongside one agent, line by line, like pair programming. In the manager view (the App grew out of the original "Agent Manager"), you hand tasks to several agents that work in the background and check in with you at milestones.

Think of a site manager: you do not watch every brick being laid; you approve the drawings, inspect at set points and sign off at the end.

Why it matters

• Agents can take on repetitive building and testing work, so ideas can be tried faster.
• Artifacts (the agent's reviewable outputs: implementation plans, code differences, screenshots, walkthroughs and browser recordings) let you check work without reading every step.
• An agent that runs commands can also do damage, so permissions matter more than in a chat.
• It does not replace testing, security review or the people who own your systems.

How to do it

1. Get approval from IT before installing it.
2. Choose a cautious permission setting. The App offers presets; "Request Review" asks before every terminal command, keeps file access inside the project folder and asks before web and connector actions.
3. Start in a practice copy, never the live system.
4. Write a short spec (see guide 15) and ask for a plan first.
5. Review the implementation plan and leave comments on it, as you would on a shared document, before the agent changes files.
6. Check the code differences, screenshots and browser recording, then test the result yourself.
7. Hand anything going live to the people who review and release code.

Example

A team wants a simple internal page listing meeting rooms and their capacity. A builder opens a practice copy of the intranet project and asks:

Goal: add a "Meeting rooms" page listing [room name],
[floor] and [capacity] from the file rooms.csv.
Rules: change only files in the pages folder. No new
outside services. Use made-up room data for testing.
First: show me an implementation plan and wait for my review.
Afterwards: give me a walkthrough, a screenshot of the page
and a browser recording of the test.

The plan proposes a new library. The builder comments "no new libraries", and the agent revises the plan, builds the page and records its test. Watching the recording, the builder sees the table does not sort by floor and asks for that. A colleague then reviews the code before release.

Common mistakes

• Choosing the most permissive setting. Fix: start with review on every action; relax it only for safe, repeated commands.
• Approving a plan without reading it. Fix: the plan is the cheapest place to catch a wrong idea; comment on it.
• Pointing agents at real data. Fix: use made-up data until the work is approved.
• Letting agents browse freely. Fix: hidden instructions in pages can steer an agent (prompt injection, see guide 13); limit which sites it may visit.

Try it this week

1. Ask IT whether Antigravity is approved and on which account (3 minutes).
2. If it is, open a practice project with the most cautious permission preset (3 minutes).
3. Ask for a plan only, for one small change, such as a new heading on a page (4 minutes).
4. Add one comment to the plan and see whether the agent changes its approach (5 minutes).

Going further

• Models available, with plan rules on the models page:

Antigravity can offer newer models than the Gemini app.

• The browser agent uses a separate browser profile; allow and deny lists control which sites it may read or click on.
• Business accounts run under Google Cloud terms, with admin logging, regional data controls and admin opt-in for other makers' models.
• Individual users can opt out of data collection in Settings.
• Rules, skills, hooks and MCP connections (see guides 11, 25 and 40) shape how agents behave.

Key words

• Agent: AI that works through a task in steps, using tools such as an editor, terminal or browser.
• Artifact: a reviewable output from the agent, such as a plan, code differences, screenshot or recording.
• Implementation plan: the agent's written proposal of what it will change, before it changes it.
• Permission preset: a saved setting that decides what the agent may do without asking.

Sources

Examples and figures in this guide are illustrative. AI tools change often: check what your organisation's approved tools offer today. Use approved tools only, and a person checks and decides.

Prefer to print or share it? Download this guide as a PDF (278 KB), or the complete AI Field Guide (2.5 MB).