Antigravity Includes a Gemini Approval Mode, but Access Is Unconfirmed

Code in Antigravity version 2.12.2 contains Vetted (beta), a permission setting that puts a Gemini-powered security agent in charge of deciding which coding-agent commands can be auto-approved. The design could reduce manual approval prompts, but the option is server-gated and has not been confirmed as available to users.

A second AI reviews proposed actions

The Vetted interface says its Gemini-powered security agent decides whether commands should be auto-approved. Recovered code and prompts identify that evaluator as Policy Guardian, giving a second model a supervisory role over actions proposed by a coding agent.

Its described checks span tool calls, downloads, software dependencies, and attempts to repeat an action that was already denied. Those are the points where an agent moves from generating an answer to trying to change a development environment or retrieve new material.

The gate leaves the practical test ahead

The setting was not activated through a live account. That leaves the rollout scope, the system’s default behavior if the evaluator cannot decide or is unavailable, and measured security performance unanswered. For teams assessing Vetted as a permission control, those details determine whether it is a convenience feature or a dependable boundary.

A decision only works if the platform enforces it

An earlier user-reported GitHub issue shows why that distinction matters. The report said a PreToolUse hook blocked ordinary tools, including file writes and command execution, but its block was not honored for invoke_subagent, which creates a sub-agent.

That reported behavior does not show how Policy Guardian works. It does put a concrete question in front of any model-based approval layer: whether a denial is enforced consistently across the routes an agent can use to act.