Antigravity for AOBies
A plain-English, no-jargon guide to Google's agentic development platform — what it is, how it's built, and how to actually get value from it on day one.
So, what is Antigravity?
Antigravity is Google's agentic development platform: a tool where an AI agent doesn't just finish your sentence — it takes a whole task, breaks it into steps, does the work (writing files, running commands, checking results in a browser), and hands you back a finished, verified result.
Two things make it click for newcomers. First, it's built on a familiar foundation — a fork of VS Code — so it looks like an ordinary code editor. Second, the agent is powered by Google's Gemini models and can see your files, run your terminal, and drive a browser, so it can carry a job end-to-end rather than stopping to ask you between every step.
Two surfaces, two ways of working
This is the one concept worth getting right. Antigravity deliberately splits into two separate windows instead of cramming everything into one. They even ship as separately-versioned apps, so don't be thrown if the version numbers differ.
Drive it yourself
The VS Code-style editor with an Agent panel on the side. Synchronous — you're in the file, working alongside one agent in real time.
- File explorer, terminal, code editor
- Tab completions and inline commands
- One agent, hands-on, immediate
- Best when you want to stay in the loop
+E
Delegate to it
A mission-control dashboard — conversations, projects, scheduled work. Asynchronous — you hand off tasks and the agents work while you do other things.
- Dispatch high-level objectives
- Run several agents in parallel (up to five)
- See status, artifacts and pending approvals
- Best for bigger jobs you can step away from
Four concepts you'll meet straight away
Agents
The workers. You give one a task; it plans, executes across files and the terminal, runs tests, fixes what breaks, and returns a result. In the Manager you can have several going at once.
Artifacts
How the agent shows its working: task lists, implementation plans, screenshots, even browser recordings. You review them to verify the logic — and can comment directly on them, like marking up a document.
Browser & terminal control
Agents can run shell commands and open a browser. That means one agent can build an app, start the dev server, load it in a browser, run tests and patch issues — without asking permission at each step.
Parallel work
The Manager lets you spawn up to five agents, each in its own workspace on a different task. You check back when ready rather than babysitting any single one.
Which surface should I use?
| If you want to… | Use | Why |
|---|---|---|
| Make a quick, precise change and watch it happen | Editor | You stay hands-on; immediate feedback in the file you're looking at. |
| Learn what the agent is doing, step by step | Editor | Synchronous pace is easier to follow while you build confidence. |
| Hand off a larger task and get on with something else | Manager | Asynchronous — the agent keeps working; you review artifacts later. |
| Run several jobs at the same time | Manager | Up to five parallel agents, each in its own workspace. |
| Keep a clear audit trail of decisions | Manager | Artifacts and approvals give you a reviewable record. |
Quick start: your first 20 minutes
Install and sign in
Download Antigravity, open it, and sign in with your Google account. It will feel familiar if you've used VS Code.
Open a real (but low-stakes) folder
Point it at a small project or a scratch folder. Don't start on anything precious while you're learning the ropes.
Start in the Editor
Open the Agent panel and ask for something concrete and small — "add a heading to this page" or "explain what this file does." Watch how it works.
Read the artifact, not just the result
When it proposes a plan or task list, actually read it. This is where you learn to trust — or correct — the agent.
Switch to the Manager (Cmd+E)
Hand off a slightly bigger task, then leave it running. Come back, review the artifacts, and approve or give feedback.
Give feedback in plain English
If something's off, comment on the artifact in normal language. You're steering, not coding the correction yourself.
How a single job actually runs
Every task, big or small, follows the same shape. Knowing it makes the agent feel predictable rather than magic.
Practitioner tips & gotchas
- 1Be specific about the outcome, not the method. "Make the pricing table mobile-friendly" beats "edit the CSS." Let the agent choose the how.
- 2Start small to build trust. Run a few low-risk tasks before handing over anything important. You're calibrating how much to delegate.
- 3Always open the artifacts. The plan and the screenshots are how you catch a wrong turn early — skipping them is how surprises happen.
- 4Use the Manager for anything you'd "kick off and walk away" from. If a task would interrupt your flow, delegate it instead.
- 5Mind the two version numbers. The Editor and the Manager update on their own schedules — different versions is normal, not a fault.
- 6Keep a scratch folder for learning. Experiment freely somewhere nothing important can break.
Want the rest of the launchpad?
FACT runs practitioner-led, applied AI training for teams who want to get genuinely useful work done with these tools. Free playbooks like this are just the warm-up.
Register your interest Back to Launchpad