Agent Manager with Google Antigravity/Antigravity Foundations
Lesson 6 / 6 · 18 min · 0 / 4 exercises
Course contents
Concept · 4 hands-on exercises

Autonomy: Permission Presets and Review Policies

Decide what an agent may do without asking: the App's four permission presets, the Plan Review Policy, your own allow, ask and deny rules, and the CLI's execution modes.

Every agent tool has to answer one question: what can the agent do without asking you? Too little autonomy and you become a click-to-approve machine. Too much and one bad instruction runs unchecked. Antigravity answers it with settings that stack: a permission preset, a plan review policy, and your own allow, ask and deny rules on top.

Four presets in the App

As of November 2026 the App offers four permission presets, and Default is selected out of the box. You change the preset in Settings > General > Global Permissions > Permission Preset, or override it for one project in Settings > Projects.

PresetTerminal sandboxTerminal commandsFilesMCP and web
Request ReviewOffAlways askProject onlyAsk
DefaultOnRun in the sandbox; ask outside itProject and temp foldersAsk
AutoOnRun in the sandbox; outside it, safe reversible commands run and risky ones askProject and temp foldersAsk
TurboOffAllowed, unrestrictedFull filesystemAllowed

The sandbox is an operating-system boundary (Seatbelt on macOS, kernel namespaces on Linux, AppContainer on Windows) that limits commands to the project and temp folders, with no network by default. Auto isn't available on business or enterprise accounts. Explore what each preset does to the same set of actions.

Explore: What runs, asks, or is blocked under each presetExercise 1 of 4
Mode

No sandbox. Every terminal command waits for you, and file access stays inside the project.

  • Read a file in the projectRuns on its own
  • Run npm test in the projectAsks you first
  • Run a safe, reversible command that can't run in the sandboxAsks you first
  • Run an irreversible command outside the sandboxAsks you first
  • Call a tool on a connected MCP serverAsks you first
  • Open a new website in the browserAsks you first
  • Anything matching your Deny rules, like reading ~/.sshBlocked
1run on their own
5need your approval
1blocked

Simplified from Antigravity's permissions docs as of November 2026. Your own Allow and Ask rules move individual actions; Deny rules block under every preset.

Move the controls and watch what changes.

Quick checkExercise 2 of 4

Switch the widget to Turbo. One action is still blocked. Why?

Plan Review Policy

The preset governs actions. The Plan Review Policy in Settings > General governs whether the agent stops after planning. Request review, shown in Settings as Always Ask, means the agent halts and asks for your approval before it acts on proposed changes; Google labels it recommended. Always proceed removes that mandatory stop. Under either policy, asking for a plan with /plan stops for your approval.

Your rules sit on top

Presets are starting points. You add rules in three lists: Allow runs without asking, Ask pauses for you, and Deny blocks immediately. Rules always take precedence over the preset, and among themselves Deny beats Ask, which beats Allow. Each rule targets an action: read_file and write_file take paths, read_url, execute_url and download_url take domains, command takes command prefixes, and mcp takes a server or a tool. The docs' own example Deny rules safeguard Git history (.git/) and SSH keys (~/.ssh). In the CLI the same rules live in a settings file:

~/.gemini/antigravity-cli/settings.json
{
  "agentMode": "accept-edits",
  "permissions": {
    "allow": ["command(npm test)", "command(npm run lint)"],
    "ask": ["command(git push)"],
    "deny": ["command(rm -rf)"]
  }
}
Fill the blanksExercise 3 of 4

Configure the CLI so file edits in the workspace apply without asking, other terminal commands still ask unless a rule allows them, and every git push waits for you.

{ "agentMode": "", "permissions": { "allow": ["command(npm test)"], "": ["command(git push)"] } }

The CLI's execution modes

ModeFile editsTerminal commands
defaultShows a diff and waits for youAsks first
accept-editsApplied automatically inside the workspaceAsks first
autoApplied automaticallyRun in the sandbox; outside it, safe reversible ones run and risky ones ask
planNone until you approve a planReads and proposes first

Shift+Tab cycles the modes, agy --mode=auto sets one for a single run, and agentMode in settings.json saves your default. In the CLI the terminal sandbox is a separate setting and it's off by default: turn it on with agy --sandbox or Sandbox Mode in /config. auto needs the sandbox and a personal account. --dangerously-skip-permissions approves every request without asking, and the name is accurate.

Earn each loosening

Start every new project on Default with plan review on. Loosen per project, not globally, and only after the agent has a track record there: a repo with a solid test suite and a run of clean reviews might move to Auto. Keep Turbo for throwaway sandboxes, never a machine that holds production credentials. Two checkpoints are worth keeping at any level: plan review before edits, which catches wrong scope while it's free, and your review of the walkthrough and diff before merge, which catches wrong outcomes while they're cheap.

You make the callExercise 4 of 4

Loosening autonomy on a real repo

You've used the Default preset on your team's API repo for two weeks. Its test suite is solid and the last eight agent runs passed review cleanly. You're tired of approving the same install prompts.

  1. What do you change?

Check your understanding
Q1. Which two checkpoints earn their cost at any autonomy level?
Q2. Under the App's Default preset, the agent wants to run a command that needs the network. What happens?
Q3. Your Plan Review Policy is Always proceed. You type /plan before a risky refactor. What happens?
· Work through the 4 remaining exercises above (stuck? each one can show you the answer).
· Score 100% on the quiz.
Vocabulary in this lesson

Plain-English definitions from the AI Coding Dictionary.