How much time do you spend each day typing another clarification to your coding agent in chat, fixing typos, and spelling out in text what you'd say out loud to a teammate next to you in three seconds? When your hands are busy inspecting architecture or logs, the keyboard becomes the slowest link in working with AI.
In the Google Antigravity 2.30 release (voice features rolled out on November 6), that barrier was removed. The environment gained full-duplex Voice Mode powered by the Gemini Multimodal Live API and the Knowledge Work interface for console-free work. I was curious to dig into the docs to see how you steer an agent by voice as an equal and what guardrails keep a project safe from slips of the tongue.
Talking as Equals: How Full-Duplex Voice Differs from Dictation
You might reasonably ask: "Didn't Antigravity already have a microphone button—what's actually new here?" It comes down to the difference between speech transcription and live conversation. Standard dictation (Ctrl+M) simply turns words into a text string you still have to proofread and send manually. The new Voice Mode is full-duplex voice communication, where Gemini Live analyzes the audio stream with under 300 ms latency, picks up intonation, and replies out loud in real time.
You are no longer dictating a draft prompt—you are having a live call with a teammate who remembers the entire chat context.
Turning it on takes one click: in the app, when the input box is empty, press the round Start Voice Mode button next to the microphone, or in the Antigravity CLI (1.5.0+) type /voice live or press F6. You get 9 voices, adjustable speech speed (0.75x–2x in the app and 0.5x–3x in the CLI), and a floating call bar when switching between chats.
Voice Safety Controls: Spoken Permissions and Mid-Flight Error Interception
By default, the agent works quietly, and saying Keep me posted prompts it to give brief summaries every 30 seconds. Thanks to instant barge-in, you don't have to wait for a step to finish. For example, during a project build you notice the agent trying to delete an important config file, and right on the fly you say: "Stop, don't touch the config, only fix the imports." As soon as it hears a stop phrase (Stop, Hold on, or Cancel that), the agent falls silent, cancels the deletion, and switches to imports. To interrupt over its speech in the console, enable /voice duplex on—just make sure to wear noise-canceling headphones so the mic doesn't pick up speaker echo.
But what happens when the agent reaches a genuinely dangerous system command?
I liked how the engineers designed Spoken Permissions. Routine requests are read aloud by the agent with a recommendation, and you reply by voice with "Allow" (once, for the session, project, or globally) or redirect the task. However, destructive operations—deletions via rm, sudo privileges, or git push --force—will never be executed on a voice command, no matter how many times you say "Allow." You can block them immediately by saying "Cancel," but to approve them you must physically click Allow on screen. For how to safeguard a project even before edits begin, see my breakdown of /plan mode in Antigravity 2.17.
Knowledge Work Interface: Agent Power Without the Terminal Window
Voice control is paired with the Knowledge Work screen for analysts, marketers, and researchers. It toggles on without a restart via Settings (Cmd+, or Ctrl+,) → General → Interface Layout → Knowledge Work. Under the hood it's the exact same agent with local files, plugins, and MCP servers (read about its role in research in our review of Antigravity expanding beyond coding), but the screen is cleared of console clutter:
- Folders instead of Git: branches and worktrees are hidden, so projects look like ordinary folders.
- Single progress line: tool calls are collapsed into one status line.
- Plain-language permissions: access prompts are translated from terminal jargon into human language.
- Hidden artifacts panel: the terminal is tucked away, and the sidebar opens only to show finished documents and spreadsheets.
Alongside this, version 2.30 added importing rules and MCP servers from Claude Code, Codex, and Cursor, plus cloud-based Remote Agents for Google AI Pro and Ultra plans.
Don't worry if talking out loud to your computer feels awkward at first and you catch yourself trying to craft textbook phrasing as if on an exam. You don't need to nail a flawless prompt on the first take: you are working with a patient partner you can pause anytime with a quick "Hold on," steer mid-sentence, mute with Alt+M (Option+M on Mac), or switch back to the terminal view in one click.

