New Updates in Google Antigravity | Episode 8
Read the Portuguese translation →
Author: Google Antigravity · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: New Updates in Google Antigravity | Episode 8 · Channel: Google Antigravity · 04:08 · Watch on YouTube
Topics: AI agents, AI developer tools, AI content generation, Local and open-source models, Health and science
Summary
Section titled “Summary”Episode 8 highlights updates focused on greater project autonomy, improved stability, and the addition of Gemini 3.8 Flash. New multi-agent commands, interactive artifacts, and sandbox improvements aim to support both everyday work and more complex tasks. The episode also covers CLI and SDK updates, including remote terminal controls and offline local AI models. It closes with examples of community projects and a reminder to check the changelog for the full list of updates.
Key points
Section titled “Key points”- Gemini 3.8 Flash is presented as a daily engineering workhorse, with the same pricing and speed as previous Flash models.
- The teamwork command coordinates agents on multi-day or frontier research challenges, while boost is intended for focused sprints on difficult tasks.
- Agents can create interactive visualizations and functional UI prototypes that run locally in a preview pane or inline in a conversation, without external dependencies.
- The updated permission system runs safe commands automatically in a project-isolated sandbox, while potentially destructive actions and external network calls still require confirmation.
- The UI now supports side-by-side terminal layouts that persist across reloads and quoting selected response text with Command-L or Control-L.
- CLI updates include persistent remote control, browser access to live terminal sessions, voice interaction, and microphone forwarding over SSH.
- The SDK supports running local AI models offline with Light RT from Google AI Edge, including open models such as Gemma 4.
- SDK updates also include OS-level terminal sandboxing and pre-built wheels for Alpine Linux container deployments.
- Featured projects include genetic research using the Alpha Genome Atlas skill, a self-generating forest, and live-tuned LED patterns on a New York City sculpture.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Gemini 3.8 Flash and multi-agent commands | The episode introduces Gemini 3.8 Flash and explains when to use the teamwork and boost multi-agent commands. |
| 01:02 | Teamwork, boost, and interactive artifacts | The video describes multi-agent research and sprint workflows, then introduces interactive visualizations and UI prototypes. |
| 02:04 | Permissions, UI, and CLI updates | Updates cover safer automatic command execution, terminal layouts, quoting agent responses, and new remote and voice CLI features. |
| 03:06 | Offline models and community projects | The SDK adds offline local models and sandboxing, followed by examples of research and creative projects built with Antigravity. |
| 04:08 | Livestream replay and closing | The episode points viewers to a livestream replay and the changelog, and invites them to share projects. |
Quotes
Section titled “Quotes”designed to be our daily engineering workhorse. — 00:30
teams of agents propose, stress test, and build solutions autonomously over hours or even days. — 01:02
anti-gravity’s agents can render interactive visualizations and functional UI prototypes. — 01:33
You can now run open models like Gemma 4 directly under local GPU. — 03:06
Concepts and entities
Section titled “Concepts and entities”Google Antigravity, Gemini 3.0 Flash, teamwork, boost, LiteRT, Gemma 4, Google DeepMind, Google Research, AlphaGenome Atlas, Leo Villareal
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Welcome back to the mission report. This episode eight. Over the past month, we work to give you deeper autonomy for your projects and better stability over your everyday workflows. And of course, we’ve added yet another new Gemini model. So, hold on to your keyboards and get ready for launch. First up, Gemini 38 Flash. If you’ve been keeping up with all of our launches, you know we dropped 38 Flash right as episode seven went live, marking our third Flash release in just six weeks. 38 Flash continue to improve
00:30 over the previous two models and designed to be our daily engineering workhorse. It is available at the same Flash pricing and speed. We’ve optimized the default anti-gravity experience to successfully complete the bulk of day-to-day tasks with [music] efficiency in mind. But for particularly complex tasks, you may want to spend some extra tokens to get the agent to reason more deeply. [music] For those, we have two specialized multi-agent Flash commands, teamwork and boost. A simple way to look at them, teamwork is for multi-day or frontier research [music] challenges, while boost is for a
01:02 focused sprint. So, with teamwork, teams of agents propose, stress [music] test, and build solutions autonomously over hours or even days. Google DeepMind and Google Research recently used teamwork to achieve breakthroughs across theoretical computer science, research mathematics, and systems engineering. Again, it uses a lot of tokens and is overkill for everyday tasks, but when you’re tackling frontier problems, give it [music] a try. And with boost, give it a shot when you hit tricky bugs, performance bottlenecks, or complex refactors. Not necessarily frontier research tasks, but tasks that really [music] make you think. Boost routes
01:33 your prompt to a focused multi-agent pipeline, pairing an autonomous implementation agent with an independent investigation agent that stress tests every change just a bit [music] more deeply. Next, sometimes terminal logs and markdown tables are not the best way for the agent to convey results or information. So now, anti-gravity’s agents can render interactive visualizations and functional UI prototypes. These artifacts [music] are interactive widgets that can run locally inside your preview pane or even in line within the conversation. And this is all done with zero external dependencies.
02:04 With our new permission system, you spend less time approving commands and more time building >> [music] >> thanks to improvements to our sandbox. Safe commands now run automatically inside a secure environment isolated to your project while potentially destructive actions and external network calls continue to require your confirmation. We also ship some great developer polish in our UI. You can now split [music] terminals side by side with layouts preserved across window reloads. Highlight any part of an agent response and press command L or control L to quote it as context for your next prompt.
02:35 Over in the CLI, you can now run remote control start to keep the CLI running as a persistent background OS service that stays alive across reboots [music] and lockouts. Or use the remote control file to connect to and drive a live terminal session from any browser with full artifact cards. You can also now speak directly to your terminal agent by pressing F5 or typing {slash} bots. And if you develop on a remote machine over SSH, the new mic serve command forwards your laptop’s microphone through an SSH tunnel straight to the remote CLI. And in the SDK, we’re introducing support
03:06 for running local AI models completely offline powered by Light RT from Google AI Edge. You can now run open models like Gemma 4 directly under local GPU. This gives you zero API costs, total privacy with code never leaving your machine, and reliable offline execution. We’ve also added OS level terminal sandboxing to keep your host machine safe during autonomous runs. And we now also provide pre-built wheels for Alpine Linux container deployments. And now some cool projects and content highlights from us and the community.
03:37 First up, in science. Researchers at Google DeepMind showed how to use the new Alpha Genome Atlas [music] skill in Antigravity to investigate genetic variants within the HBB gene. It’s a great look at how complex biological research is being accelerated [music] with agents. Over on X, a creator built an interactive self-generating forest made entirely of colorful pencil strokes powered by Gemini 3.0. And check out how artist Leo Villareal teamed up with Google Cloud and used anti-gravity to live tune LED patterns on his iconic light sculpture in New York City. And lastly, our team hosted another live
04:08 stream walking through workflows, planning strategies, and best practices for creating artifacts in anti-gravity. Check out the replay for the full episode. We love seeing what you build. So, keep tagging us in so we can keep featuring your projects. And that’s it for episode 8. As always, these are just the big highlights. So, check out the change log for the full list of updates. And be sure to subscribe or follow and we’ll see you on the next mission report.