Google Antigravity launches /swarm: whole AI agent teams powered by Gemini 4 Carbon

What's going on

On November 10, 2026, Google announced two updates for Antigravity, its development environment for agentic coding. Instead of writing every line yourself, you hand tasks to AI agents that work through the code largely on their own. First, Gemini 4 Carbon is now available in Antigravity. It's Google's most powerful model for long software projects, knowledge work, research and cybersecurity. Second, there's a preview of a feature called /swarm. Instead of a single agent, a whole team of specialized agents works at the same time.

For now, it's available to subscribers of Google AI Ultra, Google's most expensive plan. Carbon runs in the Antigravity app, the Antigravity CLI (the command-line tool) and the extensions for development environments. /swarm is rolling out now in the app and CLI. Businesses get Carbon through Gemini for Business. Researchers can apply to use Argon and Carbon free of charge for six months on Antigravity's highest usage tier.

What /swarm does differently

Google describes the evolution in three stages. A single agent handles well-defined tasks fine but breaks down once a change touches dozens of files and test suites. The next stage was subagents: a main agent dispatches helpers for individual tasks. But the bottleneck remained the main agent's context, its working memory, because it had to review, summarize and pass on every step itself.

/swarm hands your prompt to an entire family of agents:

• A coordinator spins up specialized workers that run in parallel.
• Each worker operates in its own Git worktree, a separate working copy of the same code project. That way, changes don't get in each other's way.
• Adversarial verifiers run tests and deliberately scrutinize the changes (diffs), much like a strict code reviewer.

All agents communicate via a shared append-only message board, a kind of bulletin board where posts can only be added, never overwritten. According to Google, idle agents wait without incurring token costs and spring into action as soon as someone needs their help. Tokens are the chunks of text that AI usage is billed by.

Visibility instead of a black box: the Swarm UI

So you can follow what the agents are doing, /swarm comes with its own interface in Antigravity's sidebar. In the CLI, there's a full-screen view via /swarm board or the Alt+S shortcut. It has three parts:

• Board: Watch live as the agent team weighs pros and cons.
• Roster: The member list shows active tool calls. You can open each agent's history and redirect it if needed.
• Graph: An interactive, rewindable timeline. It shows how the team got from the initial request to verified code.

Mixed teams: Carbon thinks, smaller models type

Each agent in the swarm can be pinned to its own model. Google recommends matching compute to the role. According to Google, Gemini 4 Argon remains a good choice for most demanding tasks. Carbon, on the other hand, is built to explore thoroughly, think deeply and run long verification loops, and it uses more tokens doing so. Google cites first place on CursorBench 4.0 and on Perplexity's WANDR. That makes Carbon, in Google's view, the pick for key roles: coordinator, architecture, security audit and critical review. Lighter, faster models handle the parallel file edits and routine work.

Google used /swarm internally for five weeks, what the industry calls “dogfooding.” Engineers deployed swarms for tricky multi-package migrations, for debugging across different subsystems and for architecture overhauls. But Google itself warns:

“Even with free idle time and dynamic model switching, parallel multi-agent systems consume quota much faster than a regular chat.”

What we know about Gemini 4 Carbon

According to several reports, Carbon was unveiled on October 28, 2026, about two weeks after Gemini 4 Argon. Crypto Briefing attributes the rapid pace to AI-assisted model development. Initially, Carbon was only available to select cybersecurity defenders through Google's Fairwind program. According to TechRepublic, the program has more than 650 partners worldwide, mainly government agencies, critical infrastructure operators and major tech platforms. The reason for the caution: a model that finds security vulnerabilities and checks whether they're exploitable helps defenders, but attackers too. With the release to Ultra subscribers in Antigravity, access is now much broader.

The benchmarks, according to Google and reports from Techgenyz and Crypto Briefing:

Techgenyz notes that Google measured the three security scores without safeguards in place. One detail also stands out: Crypto Briefing writes that Carbon more than doubled Argon's success rate at finding vulnerabilities. According to Techgenyz, however, that doubling refers to Google's internal tests on generating secure code, and the table values show a smaller gap for finding vulnerabilities. TechRepublic reports that Carbon ranked first overall in the Artificial Analysis index but didn't lead every category, such as coding and knowledge work. Google also reports internal safety figures: about half as many unsupported claims as Argon and almost 80 percent fewer severe misalignment flags. These numbers haven't been independently verified.

On API pricing, Techgenyz cites an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon. Later, it's set to rise to $4 and $20 respectively. When that will happen is unclear. Cached input is 95 percent cheaper.

Our take

/swarm is Google's answer to a shift sweeping the whole industry: AI coding tools are evolving from a single assistant into a coordinated team. Separate Git worktrees, a shared message board and dedicated review agents go straight at the usual problems: changes collide, the context fills up and nobody critically checks the work. Making every step visible in the interface is a smart move, because agent swarms are otherwise hard to keep under control.

What remains unclear is how well this works outside Google. The five weeks of internal use are the company's own account, and there are no independent tests of /swarm yet. Carbon's benchmarks also come mostly from Google itself. And it's expensive: you need the Ultra plan, and Google admits swarms eat through your quota fast. According to Crypto Briefing, Carbon competes with OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. If your team works on large, tangled codebases, /swarm is worth a look. For small tasks, a single agent is probably still enough.