Antigravity’s /swarm runs subagents over a message board
Subagents in Google’s Antigravity App now get a shared place to talk. A new /swarm command, added in the November 9 release listed in Google’s Antigravity changelog, coordinates multiple subagents over a message board to complete long-horizon tasks. The command ships in the desktop app’s release, and accounts get it over several days, since Google says new versions roll out gradually and may take a few days to reach all users.
TestingCatalog, which had spotted the mode earlier, describes /swarm as an agent orchestrator command that runs multiple agents in parallel, with the spawned agents using a shared message board to communicate with each other.
Running subagents side by side is not new to the product. Antigravity 2.0, announced at Google I/O 2026 as a standalone desktop app, already let users orchestrate multiple agents in parallel and shipped dynamic subagents for parallelized workflows, Google said in its I/O developer post. What /swarm adds is the board those subagents coordinate over. The release before this one had already moved notices that a subagent sent a message into the turn’s Worked for section, so conversations stay compact.
The entry sits under the Antigravity App; the same changelog tracks the CLI, the SDK and the IDE as separate surfaces, and this release lists /swarm only for the app.
Several of the release’s 11 improvements and 18 fixes touch setups beyond the default. MCP App file viewers can now save changes to local files, so a diagram or design editor plugin can edit the file it opened. A custom model gateway whose base URL ended in /v1beta used to fail, and one whose thinking model returned only reasoning before its answer reported an empty response; both are fixed. A model your Google Cloud project cannot access now returns a message explaining the access problem instead of a raw error, and Pay-As-You-Go setup suggests your Google Cloud projects as you type rather than asking for the project ID by hand. On the interface side, font size can be set to 12, 14 or 16 pixels in Settings > Appearance, plugin views can share a split view with a conversation, and a finished voice call now leaves a short summary on its transcript card.
For long jobs, the browser agent changes matter most. When it tries to download a file, the download now pauses until you approve it, so a long task that has the browser agent fetch a file waits at that step for a person. Pages the browser agent controls can no longer lock your mouse pointer or keyboard. Two fixes address runs that stalled or broke on their own: requests used to fail when a thinking model took more than a minute to start responding, and with Always Proceed set for artifact reviews, an artifact that asked for feedback hid its Proceed button and left the agent waiting.
Up next
US online businesses can publish their own agent in ChatGPT, Video
“Clearly labeled” and “separate from ChatGPT’s independent answers” is how OpenAI describes chats with its beta Business Agent, which US online businesses fill with their own prices. Would you ask a store’s own agent about a product, or stick with ChatGPT’s answer?
Related stories
- Cloud-run agents in Antigravity start reaching Ultra and Pro
- Gemini 4 Carbon reaches Agent Studio with 1M output tokens
- AI Studio now builds Google Maps apps, 100 free calls a day
- OpenAI opt-out skips personal Daybreak Blue sessions
- GitHub Copilot gets five new features at Universe’26
- Google opens Android Studio and CLI to Claude Code and Codex
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

