In Antigravity, entering /swarm lets a coordinator Agent assign cross-file software engineering tasks to parallel Agents running in separate Git worktrees. A verification Agent runs tests and reviews the diffs, while users can observe and guide the team through the Board and roster views.
Suitable tasks: Software engineering work involving multiple files, test suites, or subsystems, such as multi-package migrations, cross-subsystem troubleshooting, and architectural refactoring.
Unsuitable tasks: The article does not assess suitability. Based on its workflow, changes with a clearly defined scope that can be completed in a single conversation usually do not need parallel orchestration.
Supported model versions: Gemini 4 Carbon. The collaboration team can also assign different models to different Agents.
Supported clients, agents, or APIs: The /swarm preview in the Antigravity App and CLI. The article says Gemini 4 Carbon is available in the Antigravity App, CLI, and IDE extension.
Recommended reasoning level and parameters: Not specified.
The official entry point is /swarm in the Antigravity App or CLI. The original article does not disclose a complete, copyable task prompt, so the following only summarizes the publicly described product workflow and is not a prompt template.
Submit a task: Enter /swarm in the composer and describe the goal. The coordinator Agent organizes parallel work based on the task.
Isolate the work: The coordinator Agent starts specialized worker Agents in separate Git worktrees to reduce conflicts between code changes.
Work in parallel: Worker Agents can use different models. Google recommends using the more capable Gemini 4 Carbon for high-value coordination, architecture design, and adversarial verification. Lighter models can handle parallel file editing and routine implementation. The article also notes that Gemini 4 Argon remains suitable for most advanced engineering and knowledge work.
Share progress: Agents collaborate on a shared append-only message board. Idle Agents are suspended and wake when other Agents need their input. The company says idle waiting consumes no tokens.
Verify diffs: A dedicated verification Agent runs tests and challenges the code diffs. The article does not disclose specific test commands, acceptance criteria, or verification prompts.
Observe and guide the work: Users can view activity on the Board, see members and tool activity in the Roster, open an Agent's execution trace and provide guidance, or review the collaboration process on the Graph timeline. In the CLI, /swarm board or Alt+S opens the live panel.
Editorial suggestion example (not an official prompt): When submitting a task, state the goal, scope, and acceptance criteria, then ask the coordinator to break down the work, implement it in parallel, and arrange independent verification. This suggestion organizes the publicly described workflow. It is not a prompt published by Google.
Google describes /swarm as a multi-agent orchestration capability in which a coordinator Agent organizes parallel worker Agents and a verification Agent.
Worker Agents run in isolated Git worktrees. They collaborate through a shared append-only message board.
The Swarm UI provides Board, Roster, and Graph views for seeing live activity, guiding Agents, and reviewing the team's execution trace.
Google says that during a five-week internal trial, the feature was used for multi-package migrations, cross-subsystem root-cause investigations, and complex architectural refactoring.
The company warns that parallel multi-agent systems can consume quota much faster than standard conversations. Zero-token idle waiting and dynamic model switching do not mean the overall task is inexpensive.
As of the article's publication date, /swarm was available as a preview in the Antigravity App and CLI. Gemini 4 Carbon was available to Google AI Ultra subscribers in the Antigravity App, CLI, and IDE extension.
This is Google's account of product capabilities and an internal trial, not an independent benchmark. The article publishes no quantitative data on task success rates, duration, token usage, or code quality.
Cybersecurity audits and adversarial verification are mentioned as role use cases. This article only documents a general software engineering collaboration workflow and does not provide security attack steps.
The verification Agent runs tests and reviews diffs, but this does not replace the user's review of the requirements, test coverage, and final code.
“Zero-token idle waiting” should not be taken to mean that the entire swarm run consumes no quota. Google explicitly warns that overall quota consumption may be faster.
In an Antigravity App or CLI that supports /swarm, prepare a routine software engineering task spanning multiple files, enter /swarm, and describe the goal. Use Board, Roster, or Graph to observe task breakdown, Agent activity, and verification. Check the final test results and code diffs. Model availability, subscription requirements, and preview status may change. This article records the information in the official article published on 2026-11-10.
Gemini 4 Carbon