Google Antigravity announced Teamwork, a multi-agent orchestration framework that lets AI agents propose, stress-test, and refine solutions autonomously over extended periods. The feature, which rolled out on August 28, 2026, is designed for problems too large, uncertain, or open-ended for a single AI agent to handle alone. Unlike loosely organized multi-agent setups that can build confidently on flawed ideas, Teamwork enforces structured challenge, with agents actively looking for flaws before the campaign builds further.

The framework operates through a familiar loop: agents generate candidate solutions, stress-test them through adversarial critique, combine the strongest pieces into better candidates, and repeat with humans owning final acceptance. Google is explicit about Teamwork's limitations: it uses a lot of tokens and is overkill for everyday tasks. Developers should expect significant costs, with even strong model runs potentially burning through weekly quota mid-task.

What Are the Five Teamwork Patterns?

Teamwork is not a single fixed setup. Instead, Google offers five configurable patterns, each tailored to different problem types. When you invoke Teamwork via the /teamwork-preview slash command, Gemini analyzes your prompt and selects the most appropriate pattern automatically.

• Iterative Coding: Best for non-decomposable problems where tight feedback loops and tests are essential; agents implement, refine, and refine again.
• Distributed Coding: Ideal for decomposable software engineering tasks that fan out into parallel workstreams requiring critic review and verification.
• Long Proof: Designed for open math and science problems, including conjecture and proof search with dead ends; agents explore, falsify, solve, and verify.
• Self-Verification: Suited for deep reasoning tasks requiring rigorous checks; agents generate solutions, verify them, and revise in a loop.
• Document Review: Built for analyzing papers and document sets; agents segment content, review it, synthesize findings, and verify conclusions.

Knowing which pattern fits your problem helps you write prompts that steer Teamwork toward the right approach. For example, describing a task as "decomposable multi-module refactor with independent verification" signals Distributed Coding, while "single algorithm, refine until benchmarks pass" points toward Iterative Coding.

When Should You Actually Use Teamwork?

Google's own guidance is clear about when Teamwork makes sense and when it does not. The framework shines when your problem is open-ended or research-grade, such as proofs, novel systems, or hard optimization challenges. It also helps when a single agent keeps agreeing with itself into a dead end, or when you can leave a machine running for hours or days to complete the task.

Conversely, Teamwork is not the right choice for small pull requests, bugfixes, or UI polish. If you need predictable cost per task or want one specialist subagent for search and tests rather than a full team campaign, stick with simpler approaches. The cost and token consumption make Teamwork a tool for specific, high-value problems, not routine development work.

What Results Has Google Achieved With Teamwork?

Google has published results from using Teamwork on frontier research. In theoretical computer science and research mathematics, Google reports that Teamwork helped address seven open problems, with five papers published on arXiv. The framework also verified Knuth Cycles constructions in Lean, a formal proof assistant, and achieved a 71% score on TCSBench, Google's internal theoretical computer science benchmark, using Gemini 3.7 Flash and 3.1 Pro models.

In systems engineering, Google built a cycle-accurate out-of-order RISC-V CPU simulator that boots xv6 and maintains cycle alignment with approximately 0.71% average error. These results demonstrate Teamwork's capability on genuinely hard problems, though Google frames them as research outcomes rather than independently peer-reviewed claims.

How to Get Started With Teamwork?

• Check Your Plan: Teamwork is available to all paid Antigravity plans, not just Ultra-tier users as in earlier previews.
• Invoke the Command: Use the /teamwork-preview slash command to trigger Teamwork; Gemini will automatically select the best pattern for your prompt.
• Budget for Tokens: Ensure you have quota or credits sized for multi-agent campaigns, as Teamwork can consume significant token counts over hours or days of execution.
• Keep Your Machine Awake: Local execution requires an awake host; pair Teamwork with Remote Control if you will not sit at your desk during the campaign.
• Frame Your Problem Clearly: Write prompts that signal whether your task is decomposable, iterative, proof-based, or document-focused to help Teamwork select the right pattern.

Is Teamwork a Finished Product?

No. Google still ships Teamwork as /teamwork-preview and documents it as a research preview, with improvements to orchestration and user interface continuing over the coming weeks. The August 2026 update represents a capability and availability expansion rather than a general availability release. Community threads asking whether this marks the official launch are partly right; the command already existed in Antigravity, but this wave broadens access and adds new patterns and research results.

Builders have flagged practical gaps that matter when deciding whether to invest in Teamwork. Requests include bring-your-own-key (BYOK) support and the ability to use different models for different subjects, comparisons of Antigravity's Model Context Protocol (MCP) store surface versus other integrated development environments (IDEs), and closing the command-line interface (CLI) versus desktop capability gap. These gaps do not cancel Teamwork's research results, but they remind developers that preview orchestration and day-to-day ergonomics are still converging.

What About Windows and Linux Support?

Beyond Teamwork's core capabilities, Google is expanding platform support. The company is working on Windows Subsystem for Linux (WSL) support and better native Windows support for Antigravity, according to Rody Davis, a senior developer relations engineer at Google's Antigravity division and DeepMind.

"We are working on WSL and better native windows support. Taking some time to really get it right and work how we want," stated Rody Davis.

Rody Davis, Senior Developer Relations Engineer, Google Antigravity and DeepMind

This move is notable because Google has historically treated Windows as an afterthought in its developer tools. WSL lets developers run a Linux distribution directly inside Windows without dual-booting or using a full virtual machine, which is critical since much modern software, including build pipelines, cloud infrastructure, and AI frameworks like PyTorch and llama.cpp, is written and tested on Linux first.

GitHub Copilot, a competing AI development tool, has already added experimental WSL support to its desktop app. In a demo, Copilot's agents ran file operations, shell commands, and git validation entirely inside a Linux distribution while the developer worked from Windows. This capability lets developers leverage Linux-native tools and environments without leaving their Windows machines.

The broader context matters: Windows is quietly becoming a serious developer platform again. Microsoft has committed to WinUI as the permanent native app framework for Windows and is open-sourcing it on GitHub. Microsoft also shipped Coreutils for Windows, bringing over 75 familiar Linux command-line utilities like ls, grep, and mv natively to Windows without needing WSL at all. Ubuntu is growing faster on Windows 11 than on native Linux PCs, according to Canonical's numbers, signaling where developers are choosing to run Linux workloads.

For Antigravity users, WSL support means agents could run file operations, shell commands, and builds inside Linux instead of translating everything through Windows first. The "native windows support" part remains unclear, though WinUI is the only candidate that fits as Microsoft's permanent native app framework.