Antigravity CLI 1.5.7 Turns On Auto Mode by Default, Even for Headless agy -p

By Lesia Kapliuk · Published November 12, 2026 · Updated November 12, 2026

TL;DR: Antigravity CLI 1.5.7 (November 12, 2026) changes the default Agent Mode from Default to Auto. Per the changelog, this includes headless agy -p runs, wherever Auto mode is available. According to the CLI's own description, Auto approves file edits and safe commands and asks before irreversible actions. On our Google AI Pro account, a nested agy -p call ran in Auto without a prompt, while the same call under --mode accept-edits was auto-denied with a permission error. To turn Auto off, set Agent Mode to Default in /config.

What changed in Antigravity CLI 1.5.7

The 1.5.7 changelog entry says sessions with no saved mode, including headless -p runs, now start in Auto "wherever Auto mode is available and in Default elsewhere". Auto mode has been selectable for a while; what changed is that you no longer have to pick it. Per the 1.4.4 notes, on Windows it needs Windows 11 version 24H2 or later and a personal Google Account, so some machines stay on Default.

On our Pro account, the first launch after the update shows a banner that is not in the changelog:

The banner reads:

Auto is now the default agent mode. Press Shift+Tab to switch agent modes, or change Agent Mode in /config.

The status line under the prompt reads auto · Gemini 3.8 Flash · medium. In /config, the hint under Agent Mode describes the modes:

Mode	What the CLI says it does
`default`	standard behavior (in our /swarm test, Default stopped for every file write and command)
`accept-edits`	auto-approve file edits, prompt for commands
`auto`	auto-approve edits and safe commands, prompt on irreversible actions
`plan`	research and plan without making changes

The permissions documentation describes Auto the same way for commands outside the terminal sandbox. In the same swarm test, Auto finished with no prompts.

How to turn off Auto mode in Antigravity CLI

For one session: press Shift+Tab until the status line no longer shows auto.
Permanently: open /config and set Agent Mode to Default. This saves "agentMode": "" to settings.json.
To keep file edits behind approval even in Auto: set Tool Permission to strict in /config. As of 1.5.7, strict takes precedence over Auto and accept-edits modes.

For a single scripted run, the CLI also accepts --mode with accept-edits, auto or plan (from agy --help). It has no default value, so for Default use the saved setting above.

What Auto mode allowed in headless agy -p: our test

The changelog says headless runs start in Auto. To see what that means in practice, on November 12, 2026 we ran agy -p in CLI 1.5.7 on a Google AI Pro account with Gemini 3.8 Flash, in an empty folder, with one harmless task per run:

Task given to `agy -p`	Mode	Result
Create `probe.txt` containing one word	default (Auto)	File created, no permission prompt
Run `agy --version` in the terminal	default (Auto)	Printed `1.5.7`
Fetch the headers of https://example.com with `curl -sI` and print the first line	default (Auto)	Printed `HTTP/1.1 200 OK`
Run a nested `agy -p "Say hello in one short sentence" --model gemini-4-argon-low`	default (Auto)	Reply: "Hello! How can I help you today?"
The same nested call	`--mode accept-edits`	Auto-denied: "a tool required the 'command' permission that headless mode cannot prompt for"

The last two rows are the interesting ones. In our November 10 subagent test on CLI 1.5.4, a Flash agent running that nested call in headless mode was denied with a permission check error, even with allow rules for the command. In 1.5.7 the same call went through with no rule from us, and when we forced accept-edits, it was denied again, this time with a message that says to add a command(...) allow rule or use --dangerously-skip-permissions. That points to the mode as the difference, though we have not run the exact old setup again and we did not test a forced Default.

What this does not show:

We didn't test 1.5.5 or 1.5.6, so we can't say which release changed the default. The 1.5.7 changelog names the change; nothing else in the notes mentions headless permissions.
We see the answer, not the model that wrote it. The nested reply looks like Argon's output, but we haven't proven it came from the nested session.
We don't know whether this Windows machine has an OS-level sandbox. The 1.5.7 notes say saved allow and deny rules now apply when an admin requires the sandbox on a machine without one, so behavior there differs.
We tested harmless tasks only. We did not try destructive commands to see where Auto stops and asks.

Permission fixes in the same release

Two 1.5.7 fixes matter if your organization enforces admin review rules:

A PreToolUse hook returning auto_approve, or a wildcard command(*) allow rule, could bypass an organization's admin-enforced review. Per the changelog, commands that would leave the terminal sandbox, commands the admin policy holds for review, and file reads or writes outside the workspace under an "always ask" policy now always ask for your approval.
Saved command allow and deny rules were ignored when an admin policy required commands to run in the terminal sandbox on a machine with no OS sandbox. Your saved rules now apply, and commands not on the list still ask.

For people running scripts, another fix is relevant: the agent now notices when a command it moved to the background is waiting for input such as a password or a y/n prompt, instead of waiting silently. The release also fixes multi-second UI freezes on very long output lines, reply stalls of 20 seconds or more while tool-call groups animated, browser sign-in on Windows (the local callback now listens on IPv4 and IPv6), and a Gemini 4 Carbon error through enterprise model gateways.

How other tools handled the same switch

Anthropic announced that new Claude Code sessions on Pro, Max and Team plans start in auto mode from August 14, 2026, with pinned defaults unchanged. Antigravity CLI is making the same move about three months later. The Anthropic post does not say how headless claude -p runs are treated, so we don't compare that part. Whether an agent should ask less by default is a fair question either way; this article only records what the CLI does now.

What this means for you

Interactive users where Auto is offered (we confirmed Google AI Pro; on Windows, personal accounts on 11 24H2 or later): the agent will now edit files and run commands it judges safe without asking. If you preferred the old prompts, switch with Shift+Tab for one session or set Agent Mode to Default in /config.
Scripts and CI that call agy -p: this is the group to check. A run that used to stop at a denial may now proceed, and the permission error we saw under accept-edits is what a locked-down mode looks like in headless runs. Pin the mode you want instead of relying on the default: pass --mode explicitly, or set "agentMode" in settings.json (an empty string means Default). Try new scripts first in a throwaway folder.
Teams with admin policies: the 1.5.6 notes say Gemini Enterprise accounts keep following their organization's admin-controlled settings and are not offered Auto mode. Check your policy after updating, since the bypass fixes above ship in the same release.
Anyone using hooks: auto_approve from a PreToolUse hook no longer overrides an admin review for commands outside the sandbox or files outside the workspace.

What we don't know

Which accounts and plans count as "where Auto mode is available". We tested a personal Google AI Pro account only.
What exactly Auto treats as a "safe command" versus an irreversible action. The CLI gives the one-line description above, and we haven't mapped its boundary.
Whether agy -p behaved differently before 1.5.7 for file edits. We have only the 1.5.4 denial for nested commands.
What a forced Default does in headless mode. We tested accept-edits only, because --mode has no default value.