Google's Anti-gravity teamwork preview runs teams of AI agents in parallel for hours. Here's how the multi-agent orchestration actually works.
What is Google Anti-gravity’s teamwork preview? #
Teamwork preview is a new feature inside Google Anti-gravity, Google’s agentic development platform, that lets a single prompt spin up a team of specialized AI agents instead of one model working alone. A lead agent plans the task with you, breaks it into pieces, assigns those pieces to worker agents, and keeps an auditor or critic agent checking the output. The whole team can then run in parallel for hours on a single job, whether that’s refactoring code, localizing a website, or building a 3D model from scratch.
TL;DR #
- Teamwork preview turns a single prompt into a coordinated group of agents, with a lead agent that interviews the user, plans the work, and then delegates to specialist workers and auditors.
- The feature runs on Gemini 4 Argon , Google’s new flagship model built specifically for long, multi-step “deep work” rather than quick chat replies.
- A Google demo localized more than 100 pages of website content into17 languages in two days , a job that would normally take human translators weeks.
- The workflow uses worker and auditor agents running in parallel, with humans staying in the loop to spot-check output rather than doing the translation or coding by hand.
- A separate demo showed the same orchestration pattern used to generate a photorealistic 3D temple model in Blender from one prompt, with an orchestrator agent overseeing an adversarial loop between builder and critic agents.
- Google is positioning this as a shift in what knowledge workers actually do: people move from producing the work to reviewing it .
- Teamwork preview is currently available to Anti-gravity Ultra subscribers , tied to the Gemini 4 Argon rollout.
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How does teamwork preview actually work? #
The pattern Google showed follows a consistent structure across different use cases. A user describes a large goal in plain language. A lead agent then asks clarifying questions before any work starts, to make sure the plan matches what the person actually wants. Only after that alignment step does the lead agent break the task into a plan and spin up a team of specialized agents underneath it.
From there, the agents split into roles. Worker agents handle the actual production, whether that’s writing code, translating text, or generating 3D geometry. Auditor or critic agents check that output against a reference or standard, flag problems, and send work back for revision. An orchestrator agent sits above the whole process, coordinating who does what and resolving the loop between workers and auditors. This lets the system run unattended for long stretches, because the checking happens inside the agent team rather than waiting for a human to review each step. This is different from a typical single-agent coding assistant, which generates output and waits for the user to approve or correct it. Teamwork preview is designed to run through many iterations on its own, with humans checking in at milestones rather than every step.
What was demonstrated in the website localization example? #
Google’s clearest real-world demo of teamwork preview came from a localization project on the Anti-gravity website itself. An employee named Nick, a former technical writer who speaks multiple languages, used the feature to translate more than 100 pages of marketing and technical content into 17 languages.
The process started small. Nick picked one language, Mexican Spanish, and ran an initial pass. The first output was accurate but sounded mechanical, a common problem with machine translation that works word for word instead of capturing natural phrasing. He gave feedback directly inside Anti-gravity, and after three or four passes, the translations started reading naturally rather than literally.
Once that single language was dialed in, the lead agent used it as a template and spun up parallel worker and auditor agents to handle the remaining 16 languages at once. The system then ran autonomously for hours across hundreds of pages. The entire project, which Nick noted would normally take weeks with human translators working by hand, finished in about two days. Human reviewers still spot-checked the output in their native languages, but their job shifted from translating to verifying.
Is this the same thing as replacing human workers? #
Based on how Google framed the demos, no. The recurring message across the launch material is that Argon and teamwork preview compress the time a task takes, not eliminate the person doing it. In the localization example, native speakers still reviewed translations before anything shipped. The pitch is that review is much faster than original translation or original coding, so the same team can cover far more ground.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Google’s own internal numbers back up the “faster, not headcount-free” framing in a different context. The company said teams using Argon inside Anti-gravity increased agentic code submissions by more than 35% while cutting code rollbacks by nearly a third, which points to people submitting more work with fewer mistakes rather than people being removed from the loop entirely.
What else has teamwork preview been used for besides coding? #
Google showed the same orchestration pattern applied to a 3D modeling task. A creative producer on Google DeepMind’s prototyping team used Gemini 4 Argon inside Anti-gravity to build a photorealistic Japanese temple complex in Blender from a single prompt, run through a headless command-line interface rather than Blender’s visual editor.
The prompt referenced a “skill,” which Google describes as documentation given to the model that defines how to approach a specific type of build. The agent first called a tool to generate architectural blueprints, then used Google’s Nano Banana image model to produce four reference images based on those blueprints. The core of the build was a multi-agent adversarial loop: one set of agents rendered images inside Blender, while an orchestrator and a critic agent compared those renders against the Nano Banana reference images and pushed for revisions. The process ran overnight without supervision. The team later added animations and procedurally generated ambient sound effects in follow-up prompts.
A separate demo applied the same idea to archaeology: raw LIDAR scan data, archival photos, and historical documents from Chaco Canyon in New Mexico were dropped into Anti-gravity, converted into clean 3D geometry using a Blender integration, and turned into an interactive website with a working solar-dial feature, referencing the site’s real history as an ancient solar observatory.
Why does this matter for people who build software? #
Multi-agent orchestration isn’t a new idea, but running it reliably for hours on real production work (translating a live marketing site, refactoring large codebases) is a harder bar to clear than a one-off demo. Google’s willingness to show a full localization pipeline, including the mechanical-sounding early drafts and the manual feedback it took to fix them, is more informative than a polished highlight reel. It shows the system needs human steering at the start (the alignment interview, the first-language proof run) before it can be trusted to scale out on its own.
For developers, the practical takeaway is that teamwork preview is aimed at tasks with a clear structure and a checkable output: code that can be tested, translations that can be reviewed by a native speaker, 3D models that can be visually compared against a reference. Tasks without a clean way to verify correctness are a harder fit for this kind of autonomous, long-running agent team, because the auditor agents need something concrete to check against.
Frequently Asked Questions #
What is Google Anti-gravity?
Anti-gravity is Google’s agentic development platform, where developers (and now broader users) can describe complex tasks and have AI agents carry them out, including writing and reviewing code, rather than just chatting with a single model.
What model powers teamwork preview?
Teamwork preview runs on Gemini 4 Argon, Google’s model built for long, multi-step “deep work” tasks across coding, finance, law, and science, as opposed to quick single-turn answers.
How many agents can work on one task?
One coffee. One working app. #
You bring the idea. Remy manages the project.
The transcript doesn’t give an exact cap, but Google’s demos show a lead agent plus multiple worker and auditor agents running in parallel, all coordinated by an orchestrator agent, scaling up to cover tasks like 17 simultaneous language translations.
Do humans still need to check the output?
Yes. In Google’s own localization example, native-speaking reviewers still spot-checked translations before publishing. The stated benefit is that review takes far less time than producing the work from scratch.
Who can access teamwork preview right now?
It’s currently tied to the Gemini 4 Argon rollout inside Anti-gravity, available to Anti-gravity Ultra subscribers.