Google previews Antigravity Swarm for coordinated coding-agent teams
Google previews Antigravity Swarm for coordinated coding-agent teams
Google began rolling out a preview of Antigravity’s /swarm command on November 10, adding a shared coordination system for multiple agents working on a task. The announcement accompanied Gemini 4 Carbon’s arrival in Antigravity, but the two features have different access rules: Swarm is documented for Google AI Pro and Ultra customers, while the Carbon rollout in the app, command line and IDE extensions is for Ultra subscribers. Google announcement Swarm documentation
The operating documentation adds detail. A user’s existing conversation becomes the coordinator, and workers begin with separate conversational context and public task briefs. The coordinator usually assigns one verifier for every two to four workers. Idle participants wait without consuming tokens and wake when mentioned; a monitor can intervene if the entire group goes quiet before finishing. Swarm documentation
One implementation detail deserves care. The launch post describes workers in isolated Git worktrees, but the documentation says workers share the workspace by default and can receive separate working copies when simultaneous editing or builds require it. Teams should check the configuration of a specific run rather than assume that all parallel edits are automatically isolated. Google announcement Swarm documentation
Google also warns that parallel agents consume quota substantially faster than a standard conversation. Assigning different models to different roles is supported, giving users a way to reserve heavier reasoning for coordination or verification while using other models for implementation. Those controls do not establish a guaranteed cost saving. Google announcement
The release makes coordination itself a product feature. For work that separates cleanly into packages, tests or research questions, parallel execution could shorten waiting time. It can also multiply the amount of output a team must review and the opportunities for agents to act on inconsistent assumptions.
The practical question is whether visible communication and dedicated verification produce better completed changes for the additional quota. Google’s internal-use examples are evidence of company adoption, not an independent comparison with a single-agent workflow. Swarm remains a preview, and its value will need to be measured on the repositories and review standards of the teams using it. Sources rechecked November 15, 2026. Original announcement dates retained.
Sources and citations
Revision history
Published .