One huge R script that’s hard to edit
Each stage can be developed and re-run independently; the figures module is reused in a second paper within the same week.
Track 13 · Advanced
13.2.4
Antigravity
Agent-first dev platform — orchestrate many AI agents at once
1Overview
Google's entry in the agent-first coding space — unlike most competitors, it needs a Google account and runs on Google's own infrastructure rather than a tool you install standalone.
Antigravity is Google's agent-first development platform: instead of one assistant in a chat box, it gives AI agents their own space to work. A Manager Surface lets you spawn, orchestrate, and watch multiple agents running in parallel, and an Editor View is a full AI IDE. It is model-agnostic — Gemini 3 Pro, Claude Sonnet 4.5, and GPT-OSS all run inside it. → Best for bigger, multi-step jobs you hand off and check on later.
1.1Before you start
AI coding assistants— Reading AI coding assistants shows how agents edit and run code within your existing project, preparing you for Antigravity.A repo you can push to— Use GitHub to manage repositories, pull requests, issues and actions—where AI agents turn issues into completed pull requests, so start there.
1.2After this chapter you can
→Spawn and orchestrate multiple agents at once in the Manager Surface
→Schedule background or overnight agent work
→Switch between Gemini, Claude, and GPT-OSS inside one platform
1.3Best for
Concrete multi-step jobs with a clear finish line — a refactor, a new tool, an overnight task — not a single quick question.
1.4Watch out
Needs a Google account and runs on Google's platform; the free tier is capped (~20 agent requests/day) and exhausts in a real session.
1.5Free vs paid
Free: ~20 agent requests/day on Gemini Flash. AI Pro $20/mo (~100/day); AI Ultra $100–$250/mo for high-volume use.
2Lessons 5
2.1 Install the Antigravity CLI
The Antigravity CLI is a terminal‑first surface for working with AI agents.
You will have the agy binary installed and reachable from your shell.
- Open a macOS or Linux terminal (or Windows PowerShell).
- Run the fast‑path install command:
curl -fsSL https://antigravity.google/cli/install.sh | bashfor macOS/Linux, orpowershell irm https://antigravity.google/cli/install.ps1 | iexfor Windows. - Verify that the installer placed the binary in the default directory (
~/.local/bin/agyon macOS/Linux,C:\Users\on Windows).\AppData\Local\agy\bin - Add the binary directory to your PATH if it is not already there (e.g., add
export PATH="~/.local/bin:$PATH"to~/.bashrcand runsource ~/.bashrc).
- You'll see
agy --versionprints the CLI version without error, confirming the command is found. - Takeaway Ensuring a tool’s executable is on your PATH is essential for reliable command‑line usage.
2.2 Launch the Antigravity TUI inside a project folder
The Terminal User Interface (TUI) lets you interact with agents via a full‑screen console.
You will start the Antigravity TUI and complete its first‑launch interactive setup.
- Navigate in your terminal to an existing code folder you want the agent to work on (
cd /path/to/your/project). - Run
agyto launch the TUI. - When prompted, choose a color scheme (e.g., Dark).
- Select a rendering mode (Alt‑Screen or Inline) as instructed.
- Confirm the workspace trust prompt so the agent can index the folder.
- You'll see A full‑screen interface appears with panels for conversations, file search, and a prompt bar at the bottom.
- Takeaway The TUI’s initial configuration tailors the UI to your environment and grants the agent access to your codebase.
2.3 Create a new Antigravity Project
A Project defines the folder boundaries and settings that agents can operate within.
You will have a named project that includes one or more local folders.
- In the TUI, click the “+” icon in the left sidebar to open the New Project dialog.
- Press the button labeled “New Project”.
- Click “Add Folder” and select the folder you opened in the terminal (or any other folder you wish to include).
- Press “Create” to finalize the project.
- Optionally, open the project settings panel and note that each project maintains isolated security policies.
- You'll see The project appears in the sidebar with its name, and the agent shows a ready status for that project.
- Takeaway Projects isolate agents’ access, allowing safe multi‑repo orchestration.
2.4 Spawn an agent to add a new feature
Agents are autonomous AI workers that can read code, propose changes, and execute tasks.
You will start an agent in your project and give it a concrete coding instruction.
- Select the project you created from the sidebar.
- In the chat input at the bottom of the TUI, type:
Help me add a new feature that prints "Hello Antigravity"and press Enter. - When the mode selection modal appears, choose Local Mode so the agent works directly in your active folders.
- Watch the agent reason about the task and propose a plan in the conversation pane.
- Approve the plan by typing
/goalto let the agent run until completion.
- You'll see The agent creates or modifies a file with a function that prints the requested message, and the console shows a success confirmation.
- Takeaway Using slash commands like
/goallets you control how much autonomy an agent has during execution.
2.5 Schedule a recurring web‑search task
Scheduled Tasks let agents run instructions at future times or on a repeat interval.
You will set up an agent to perform a web search every hour and store the results as artifacts.
- In the TUI conversation input, type:
/schedule 1h Write a summary of the latest AI research papersand press Enter. - Confirm the scheduling prompt that appears, selecting the default recurrence of one hour.
- Switch to the Artifacts tab in the side panel to view where the agent will store its output.
- Wait for the first execution (or trigger it manually with
/run now). - After execution, open the generated artifact file and verify it contains a summary.
- You'll see An artifact appears in the project’s artifact list with a timestamped filename, containing the requested summary.
- Takeaway Scheduled tasks enable hands‑off automation, turning agents into recurring assistants for ongoing knowledge work.
3You’ll know it worked 7 checkable outcomes in this chapter
- ✓All figure outputs are unchanged when running run_all.R
- ✓All tracked events appear in the network tab when clicking through flows
- ✓Normalisation function passes all generated pytest cases
- ✓A summary email is sent each night listing any test failures
- ✓Function get_sequence_by_accession exists and returns the correct sequence when called with an accession ID
- ✓The terminal displays "Welcome to Antigravity CLI!" followed by a colour-scheme picker
- ✓The agent (shown as AGY) outputs a diff preview of the proposed code changes
4FAQ, Tips & How-to 27
one problem, one solution, one action
Research & data tools1
Dashboards & analytics1
Want a complete funnel view without hand‑coding tracking
The founder has a complete funnel view within one day of setup: no manual instrumentation pass, and no manual click-through either, because the agent verified its own work in a live browser.
Internal tools & ops3
Typed modules improve maintainability while tests catch regressions early in data cleaning.
Nightly test runs
You’ll catch regressions early without manual effort, keeping the codebase stable.
Need a sequence for an accession ID
Enables rapid reuse of existing data without manual download, streamlining downstream analyses.
Want to choose a project and keep the default Gemini model
Choosing a project and leaving the model on Gemini 3.5 Flash prepares the agent with a capable default LLM
Unsure which part of the code to tweak
Stating exactly which code to modify and its location helps the agent produce targeted edits instead of vague improvements
The Editor View shows the agent's planned steps, code execution, and automatic fixes, giving you visibility into its self-correction process
Too many small edits exceed free limits
Grouping related changes into a single task reduces the number of API calls, keeping you under the basic weekly rate limits of the free Individual plan
Need to modify just one branch directly
Running an agent locally lets it modify the selected branch directly, suitable for small trusted changes
Can't decide on terminal colours
You can launch the Antigravity command-line interface and immediately choose a colour scheme for your terminal
Need to request a code change in plain English
The CLI forwards your natural-language request to an agent that replies with a ready-to-review code diff
Running many agents in parallel consumes the same weekly rate-limit, so total throughput increases but cost does not decrease
Too many AI agents get blocked by rate limits
Upgrading to the Google AI Pro plan increases the rate-limit pool, allowing more concurrent agents without hitting the free-plan ceiling
Want a task to keep running after you log off
You can start work for an agent and let it finish on its own, freeing you to do other things
Need nightly test runs with failure list
You can ask an agent to pull commits, run tests nightly, and summarize failures for morning review
When you want to give feedback on generated docs
Leaving comments on the generated artifacts lets you guide the agent to fix issues
What is Google Antigravity?
Google Antigravity is an agent-first development platform launched in November 2025. Instead of just suggesting code like a chatbot, it autonomously plans, executes, and verifies complex tasks using AI agents that can operate your editor, terminal, and browser simultaneously. Think of it less like smart autocomplete and more like delegating a whole project to an AI assistant that works through steps on its own.
What is the Browser Subagent and what does it do?
The Browser Subagent opens a real Chrome instance the AI controls directly — it can navigate to your web app, click buttons, fill forms, and report what it finds or what breaks. This lets an agent test a website end-to-end without you doing it manually, a capability several competitors lack at this price point.
What are the free-tier quota limits, and have they changed?
The free tier launched with a generous daily agent-request allotment in November 2025 and was cut sharply within weeks, followed by several further quota reductions through early 2026 that caused some Pro subscribers to hit multi-day lockouts. Basic editor features like tab completion remain effectively unlimited, but autonomous agent usage is tightly capped — verify the current quota before relying on it.
How is it different from a chatbot like Gemini or ChatGPT?
A chatbot answers questions in a conversation window and leaves the actual work to you. Antigravity's agents take actions directly — writing files, running code in the terminal, opening a browser, clicking buttons — and keep going through multi-step tasks without you copy-pasting each response. The key difference is autonomy: the agent does, not just says.
Is it free? What are the pricing plans?
There is a free tier ($0/month) during public preview with core agent access and a Gemini model included, but it comes with weekly usage quotas. Paid plans go through Google AI subscriptions (Pro around $20/month, with higher Ultra tiers). Free limits have been cut several times since launch, so heavy use will likely require a paid plan.
Which AI models does it use? Am I locked into Gemini?
No, it is model-agnostic. Antigravity supports Google's Gemini models (the default), Anthropic's Claude, and OpenAI's open-weight GPT-OSS models, and you can pick different models for different tasks. Using a non-Google model may require your own API key from that provider, which adds cost.
Do I need a Google account to use it?
Yes, you sign in with a Google account. Personal Google accounts work for the free and Pro tiers; Google Workspace accounts are used for enterprise/team plans, subject to admin settings and geographic availability. There is no anonymous or guest access.
What platforms does it run on — Mac, Windows, Linux?
Antigravity runs on all three: 64-bit Windows 10 or later, macOS 12 or later, and 64-bit Linux. The CLI was noted as preview-quality on Linux as of mid-2026, meaning it may be less polished there.
What is the Manager Surface / Mission Control?
The Manager Surface (called Mission Control in version 2.0) is a dedicated panel where you spawn, watch, and control multiple agents running in parallel across different projects. Each agent shows its status and produces Artifacts — verifiable outputs like task lists, screenshots, or recordings — so you can check what it did without reading thousands of lines of code.
Can I run multiple agents at the same time or in the background?
Yes — this is a core feature. You can launch several agents simultaneously from the Manager Surface, each working on a different task asynchronously so they run while you do other things. Version 2.0 added scheduled background tasks so agents can also run on a timer without you present.
The same set on /recipes, filtered by tool and role.
5Videos 3
Stop Coding, Start Orchestrating: The Ultimate Antigravity 2.0 Guide
Shows why Antigravity differs from a normal IDE: you orchestrate a team of agents rather than autocomplete one file.
How to Install & Use Google Antigravity IDE 2.0 (Gemini 3.5 Flash)
The shortest path from 'never opened it' to a working setup — pair it with the official guide above.
Google Antigravity 2.0 Beginner's Guide
Straight from Google — the authoritative starting point before any third-party take.
6FAQ 10
What is Google Antigravity?
Google Antigravity is an agent-first development platform launched in November 2025. Instead of just suggesting code like a chatbot, it autonomously plans, executes, and verifies complex tasks using AI agents that can operate your editor, terminal, and browser simultaneously. Think of it less like smart autocomplete and more like delegating a whole project to an AI assistant that works through steps on its own.
What is the Browser Subagent and what does it do?
The Browser Subagent opens a real Chrome instance the AI controls directly — it can navigate to your web app, click buttons, fill forms, and report what it finds or what breaks. This lets an agent test a website end-to-end without you doing it manually, a capability several competitors lack at this price point.
What are the free-tier quota limits, and have they changed?
The free tier launched with a generous daily agent-request allotment in November 2025 and was cut sharply within weeks, followed by several further quota reductions through early 2026 that caused some Pro subscribers to hit multi-day lockouts. Basic editor features like tab completion remain effectively unlimited, but autonomous agent usage is tightly capped — verify the current quota before relying on it.
How is it different from a chatbot like Gemini or ChatGPT?
A chatbot answers questions in a conversation window and leaves the actual work to you. Antigravity's agents take actions directly — writing files, running code in the terminal, opening a browser, clicking buttons — and keep going through multi-step tasks without you copy-pasting each response. The key difference is autonomy: the agent does, not just says.
Is it free? What are the pricing plans?
There is a free tier ($0/month) during public preview with core agent access and a Gemini model included, but it comes with weekly usage quotas. Paid plans go through Google AI subscriptions (Pro around $20/month, with higher Ultra tiers). Free limits have been cut several times since launch, so heavy use will likely require a paid plan.
Which AI models does it use? Am I locked into Gemini?
No, it is model-agnostic. Antigravity supports Google's Gemini models (the default), Anthropic's Claude, and OpenAI's open-weight GPT-OSS models, and you can pick different models for different tasks. Using a non-Google model may require your own API key from that provider, which adds cost.
Do I need a Google account to use it?
Yes, you sign in with a Google account. Personal Google accounts work for the free and Pro tiers; Google Workspace accounts are used for enterprise/team plans, subject to admin settings and geographic availability. There is no anonymous or guest access.
What platforms does it run on — Mac, Windows, Linux?
Antigravity runs on all three: 64-bit Windows 10 or later, macOS 12 or later, and 64-bit Linux. The CLI was noted as preview-quality on Linux as of mid-2026, meaning it may be less polished there.
What is the Manager Surface / Mission Control?
The Manager Surface (called Mission Control in version 2.0) is a dedicated panel where you spawn, watch, and control multiple agents running in parallel across different projects. Each agent shows its status and produces Artifacts — verifiable outputs like task lists, screenshots, or recordings — so you can check what it did without reading thousands of lines of code.
Can I run multiple agents at the same time or in the background?
Yes — this is a core feature. You can launch several agents simultaneously from the Manager Surface, each working on a different task asynchronously so they run while you do other things. Version 2.0 added scheduled background tasks so agents can also run on a timer without you present.
7Glossary 18 terms
Show the 18 terms
Antigravity
Antigravity- Google's standalone desktop application for AI-assisted software development, where one or more AI agents autonomously plan, write, and test code on your behalf.
Antigravity CLI- A command-line version of Antigravity that lets you run agents directly from your terminal without opening the desktop app.
Antigravity SDK- A set of developer tools that lets you embed Antigravity's agent capabilities inside your own programs or infrastructure.
Agent- An AI process that Antigravity launches to carry out a task — it can read files, run code, browse the web, and make changes autonomously.
Manager Surface- The part of the Antigravity interface where you can launch, watch, and control several agents running different tasks at the same time.
Editor View- The traditional code-editor side of Antigravity, which supports AI tab-completions and inline commands for developers who prefer a hands-on coding style.
Artifacts- Structured outputs an agent produces — such as task lists, implementation plans, code diffs, screenshots, and walkthroughs — so you can review what it did before accepting the changes.
Skills- Packaged instruction files that teach an agent how to handle a specific kind of task; you can add them globally or per project.
Review-Driven Development- An Antigravity mode where the agent works independently but pauses at important steps so you can approve or reject significant changes before they are applied.
Plan mode- An agent mode where the agent first produces a written implementation plan for you to inspect and approve before it starts making any changes to your code.
Fast mode- An agent mode for small, single-step tasks where the agent acts immediately without stopping to show a plan first.
Project- A folder (or set of folders) you open in Antigravity that defines the workspace boundary — the files, tools, and permissions the agent is allowed to use.
Conversation- A message thread inside a project where you type instructions to the agent and see its replies; one project can contain many conversations.
MCP- Model Context Protocol — a standard way to connect external tools (databases, APIs, services) to an agent so it can use them during a task.
/browser- A slash command that tells the agent to open a real web browser and carry out a web-based action you describe.
/schedule- A slash command that sets up a task to run automatically at a fixed time or on a repeating schedule, without you having to start it manually.
Gemini 3.5 Flash- The default AI model powering Antigravity agents — Google describes it as four times faster than frontier models while outperforming them on most coding and agentic benchmarks.
SKILL.md- The required file inside a Skills package that contains both YAML metadata (so the agent knows when to load the skill) and Markdown instructions (telling it what to do).
8See also
💬 Discuss this chapter
Ask, share, or report — over on the Heidelberg AI community forum.