Working with Antigravity, Tested

A guide to Google’s assistant for people who direct it rather than program: Antigravity and its command-line app, the Gemini app beside it, the rules file it reads, and which habits held up on a small folder of files

What works when you direct Google’s Antigravity to do real work: which Google assistant you are using now that Gemini CLI has gone, the instruction file it reads, four habits measured on the same made-up folder, the weekly limits, and the setting that decides whether your work trains Google’s models. Last checked November 2026.
Author

Behzad Samadi

Published

November 12, 2026

A paper-craft garden committee table: an open laptop beside a glass cold frame holding a folder of notes with sheets floating up inside it, a padlock and tag on its lid, seedlings, gloves, a mug, and the five-dot AI glyph above. Title: Working with Antigravity, Tested.

Last checked 12 November 2026, on Gemini 4 Argon (Medium), in Antigravity CLI 1.5.6.

This guide is for someone who uses an assistant to get things built or sorted out, not to write code themselves. It covers Google’s assistant as it stands in November 2026: Antigravity, the app and the command-line version that work on the files in a folder, and, more briefly, the Gemini app for chatting in a browser or on a phone.

If you used Gemini CLI, it has gone for personal accounts. Google moved it into Antigravity, and on 18 June 2026 Gemini CLI stopped answering Google AI Pro and Ultra subscribers and free users. Antigravity’s command-line app, agy, reads the same rule files, so most of what you set up carries over.

The practices in this guide were tried on the same small folder of made-up files as the other guides in this series: a garden club’s seven sets of committee minutes, with untidy names, and a members list with four traps in it. Each task was asked plainly and then with the habit being tested, three times each, each run in a fresh copy of the folder. The counts are small, so the guide says “in three runs out of three”, not “always”.

If you have not yet let an assistant work on a folder, start with Let It Open the Folder. For Antigravity beside other coding agents, see One Repository, Three Agents.

NoteIn short
  1. Antigravity got the facts right on every task it answered, asked plainly or carefully: the open actions, the fee total with its traps, the meeting held a day late, and a figure that was not there.
  2. What it does not do on its own is follow your house style. Asked plainly for a notice to members, it wrote too much, ended with an exclamation mark and signed it with a committee member’s name, in three runs out of three. Give rules like those to it in writing.
  3. The habit that made the most difference is the instruction file: the house rules saved as AGENTS.md in the folder were kept in three runs out of three. “Only from these files” mattered too: asked plainly, it twice stopped with no answer after trying to open a file it was not allowed to.
  4. Keep it on a short lead. Antigravity’s careful auto mode was not offered on the test account; a short list of allowed commands did the job without approving everything. And switch off Enable Telemetry if you do not want your work used to improve Google’s products.

Which one you are using

What it is Works on your files Comes with
Gemini app Chat at gemini.google.com and in the phone apps Files you upload Every Google account; more on Google AI plans
Antigravity app A desktop “command center” for agents working in your project folders Folders you open in it Personal Google accounts, free tier included
Antigravity CLI (agy) The same agent, typed in a terminal The folder you start it in The same accounts
Antigravity IDE and editor extensions Antigravity inside a code editor Your project The same accounts

You sign in with a personal Google account, aged 18 or over, in a listed country. A Google Workspace account works only if your administrator has switched Antigravity on. Which models you are offered depends on your plan; see The moving parts. Run agy models to see yours: on the test account the paid plan’s models were not always offered, so check before a long job.

This guide was tested in Antigravity CLI, because a script can run it the same way each time and keep what it wrote. The Antigravity app shares the agent; the Gemini app was not tested, and what follows about it comes from Google’s own pages.

How the tests were run

Each run was one command, agy -p "<prompt>", in a fresh folder, asking for Gemini 4 Argon at its medium thinking level (--model gemini-4-argon-medium, the default on the test account’s plan that day; no separate effort setting). Antigravity’s JSON reply does not name the model; the quota that fell during the runs was the Argon one, and no other. Google’s auto mode, which approves edits and safe commands, was “not available for this account”, and without it a run that cannot ask for permission is refused every command. So the runs used --mode accept-edits --sandbox with a short list of allowed commands passed for that session only: reading and listing files, mv and cp, checksums, python3; refused were deleting, network tools, any command outside the sandbox, web pages and connections to other services (MCP). No connections or plugins were set up on the account. Antigravity’s own web search tool was available; the answers cite only the folder’s files. The only instruction file in any run was the AGENTS.md of the instruction-file test; no global rules file was set. Permission rules already saved in the app’s own settings still applied, as they would for anyone.

Giving it context

Files. In the Antigravity app or CLI, the agent opens what it needs in the folder; type @ to point at a file. In the Gemini app you upload them.

The instruction file. A plain text file called AGENTS.md or GEMINI.md in the folder. Google: “Create a GEMINI.md or AGENTS.md file at your workspace root to outline specific directory standards, styling paradigms, test command parameters, and deprecation warnings.” Both are always read; rules that apply everywhere go in ~/.gemini/AGENTS.md or ~/.gemini/GEMINI.md in your home folder. All the rules together have a budget of 20,000 tokens; past that, the largest files are only pointed to, and read when needed.

The house rules saved as AGENTS.md in a brand-new folder were read at once: asked what rules it had, Antigravity quoted them in full. No trust step was needed, unlike some other assistants.

Gems (the Gemini app’s saved instructions and files) are going: “Starting in November 2026, Gems will go away on personal Google accounts and automatically transition to skills.” Memory in the Gemini app was not checked for this guide.

What the instruction file did

The task: Write the notice to members about the work day, from the committee’s notes. Save it as notice.md. The club has six house rules for anything sent to members:

- Dates in full, with the weekday and the year: Saturday 14 November 2026.
- Times on the 24-hour clock: 09:30, not 9:30 am.
- British spelling.
- No exclamation marks.
- Under 120 words, heading and signature included.
- Signed "The committee", with no personal names.

Run three times without the rules and three times with them saved as AGENTS.md. Marked out of 10: four facts (the date, the times, gloves, soup) and the six rules.

Run 1 Run 2 Run 3
No instruction file 7 7 7
Rules in AGENTS.md 10 10 10

Every notice had the four facts, the full date and 24-hour times. Without the file, all three ran to 153 to 176 words, ended “We look forward to seeing you there!”, and were signed “Ibrahim Solheim, On behalf of the … Committee”; one also listed all five committee members by name. With the file, every rule held, at 79 to 107 words, signed “The committee”.

Practices that held up

Tested on 11 and 12 November 2026, in Antigravity CLI 1.5.6, on Gemini 4 Argon (Medium). Each was run three times plainly and three times with the practice, each in a fresh copy of the folder. No run wrote or changed a file it was not asked to.

Brief it, do not just ask it

Plain: What do we still need to do?

With the practice: Goal: a list of the committee’s open actions, so nothing is forgotten before the garden closes for winter. Read every file in notes/. An action is something a named person, or the committee, agreed to do. It is open unless a later note says it was done. If an action was carried over, list it once, with its latest date. Write actions.md with one table: action, who, due date, the note it was last mentioned in. Under the table, list anything you were unsure about and why. Do not change any other file.

Marks out of 9 Run 1 Run 2 Run 3
Plain 9 9 9
Brief 9 9 9

Both found the five open actions with the right people and dates, in three runs out of three. The plain answers, about 340 to 420 words, also mixed in jobs nobody had agreed to: “provide soup”, deal with the duplicate member row, reassign the plot of the member who moved away, and, in two runs, pay the year’s water bill. The brief gave exactly the five, in a file, and put the doubtful cases under the table: who pays the roofer once the $80 is explained, who brings the soup, whether the volunteer chasing the city counts as committee.

Plan first when it is going to change files

With the practice, first message: I want the notes folder tidied: every note named for the date of its meeting, as YYYY-MM-DD-committee-meeting.md, and a notes/index.md that lists each note with one line on what was decided. The text of the notes must not change. Do not change anything yet. Tell me your plan: what you will rename to what, how you will work out each date, and how you will check afterwards that no text was lost. Then stop.

Second message, same conversation: Go ahead with that plan. When you finish, tell me what you checked.

The plain version was Tidy up the notes folder: consistent file names, and an index.

Run 1 Run 2 Run 3 Said how it knew nothing was lost
Plain passed passed passed no, in three runs out of three
Plan first passed passed passed yes, in three runs out of three

All six renamed the notes by the date of the meeting, kept the text, wrote an index, and noticed that the file called “25 aug” was the meeting held on the 26th. The plain runs said nothing about how they knew the text was intact. The planned runs stopped as asked, then afterwards gave a fingerprint (a SHA-256 checksum) and the size of every note before and after.

Ask for the evidence, not just the answer

Plain: How much have we collected in fees this year?

With the practice: How much have we collected in fees this year? Before you answer, look through members.csv for anything that would make a simple sum wrong: duplicates, blanks, refunds, odd formats. List each one you find and say how you treated it. Then give the total, the number of members, and how you checked the total.

The right answer is $855.00; a plain sum of the column gives $900.00. Marks out of 5: the total, and each of the duplicate member, the blank fee and the refund named.

Run 1 Run 2 Run 3
Plain 5 5 5
With the practice 5 5 5

The plain question already got $855.00, named all three traps and showed the breakdown, in three runs out of three. The careful one added what it asked for: the number of members, explained, and two checks of the total, one adding up by fee and one working down from the raw $900.00.

Say “only from these files”

With the practice: Answer only from the files in this folder. If the answer is not in them, say so, and say what the files do contain on the subject. What did the club pay for water in 2025?

The 2025 water bill is not in the files; the 2026 ones are, and so is a 2025 electricity bill.

Run 1 Run 2 Run 3
Plain right no answer no answer
Only from these files right right right

Every answer it gave said the 2025 figure is not there, gave the 2026 figures and pointed out that the 2025 bill in the notes is for electricity. None made up a number. But two of the plain runs tried to open a file the test’s allow list did not cover, were refused, and stopped after a few seconds with nothing to show. The test folder was all there was to read, so the likeliest reading is that Antigravity went looking outside it. With the careful wording it stayed in the folder in three runs out of three.

What did not hold up

“Run agy inspect first.” Repeated in a troubleshooting guide. Antigravity CLI has no such command; it belongs to another assistant. Use agy -p /model and agy -p /usage to see the model and the limits, and agy mcp list to see connections.

The other repeated tips were not tested: the paid plan’s models disappeared from the test account before they could be.

What the tests do not show

  • Three runs each, on one model on one day. Gemini 4 Argon (Medium). Three earlier runs of the plain open-actions question on Gemini 3.8 Flash, before the paid plan reached Antigravity, scored the same 9 out of 9 and are kept in the research folder.
  • The tasks were easy for this model where facts were concerned. The instruction file and “only from these files” changed the result; the brief, the plan and the request for evidence changed what Antigravity showed you about its work.
  • How the runs were allowed to act. A short allow list, not the auto mode a reader with a different account may have. With more freedom, the two plain water runs might have answered, or looked further afield.
  • The Gemini app was not tested, and neither was the Antigravity app.

Where your data goes

What the agent reads in your folder, and what you type, goes to Google to be processed.

Antigravity. Its terms: Google records “your user data, interaction data … and any feedback you provide”, and uses them “to evaluate, develop, and improve Google and Alphabet research, products, services and machine learning technologies”. “Google employees and contractors may access, view, review and use Interactions.” The switch is Enable Telemetry: in the app under Settings > Account, in the CLI under /config. Even with it off, “Google still uses your data to provide the service, process any feedback and help keep Antigravity safe.” It is always off for business accounts.

The Gemini app. With Keep Activity on (at myactivity.google.com/product/gemini), Google uses your chats to “develop, and improve its services (including training generative AI models)”. Activity is deleted after 18 months unless you choose 3 or 36 months. With Keep Activity off, chats are still kept for 72 hours and are not used for training unless you send feedback. “A subset of chats are reviewed by human reviewers”, and those are kept for up to three years even if you delete your activity. Temporary Chat is not used for training.

One more rule worth knowing. Google’s FAQ says agy -p “is intended for direct scripting and CI workflows”, and that wrapping it inside other companies’ coding agents “is not supported under consumer plans”. Its terms call using third-party tools to reach Antigravity a breach that can lead to your account being closed.

Settings worth changing:

  1. Enable Telemetry, off, in Antigravity.
  2. Keep Activity in the Gemini app, off, or use Temporary Chat for private things.
  3. On folders you care about, leave Antigravity asking before commands, or allow a short list of them in its settings; do not use --dangerously-skip-permissions, which approves everything.
  4. Give it a copy of the files it needs, not your whole Documents folder.

The moving parts

As read on Google’s pages on 12 November 2026. No prices are given here: the Google One plans page showed none.

Free & Google AI Plus Google AI Pro Google AI Ultra
Antigravity limit “A meaningful quota, refreshed weekly” refreshed every five hours, up to a weekly limit the highest, refreshed every five hours
Gemini models in Antigravity Gemini 3.8 Flash adds Gemini 4 Argon (paid, not trial) adds Gemini 4 Carbon
Other models Claude Sonnet and Opus 4.6, GPT-OSS-120b Claude 4.6 and 5.5, GPT-OSS-120b Claude 5.5, GPT-OSS-120b
Extra use past the limit no AI credits, if you allow it AI credits, if you allow it
  • Claude 4.6 and GPT-OSS-120b are marked “Will be removed on November 18, 2026”.
  • What the tests used. On Google AI Pro, agy -p /usage showed a five-hour limit and a weekly limit for Gemini models, and a separate weekly limit for Gemini 4 Argon. The thirty runs took the Argon weekly limit from 99% to 81%, about 0.6% a run, and did not move the five-hour figure. Earlier, when the account was offered only the free tier’s models, three runs on Gemini 3.8 Flash took that weekly limit from 89% to 72%, and two of the three, run at once, hit “You have exhausted your capacity on this model” and retried. Antigravity reports usage as a share of the limit, not in money.
  • Antigravity CLI 1.5.6. Default model on the test account’s Pro plan: Gemini 4 Argon (Medium); when only the free tier showed, Gemini 3.8 Flash (High). auto mode was “not available for this account” either way.
  • Gemini CLI stopped serving Google AI Pro, Ultra and free individual users on 18 June 2026; enterprise and paid API keys continue.

What changed

  • 12 November 2026. First published.

Sources

Google’s pages, read on 12 November 2026. The research notes, the test log and each run’s output are in the accompanying folder. The made-up folder, prompts and answer key shared by all the assistant guides are in the common folder.