Inside Google's Antigravity Swarm: three-agent test exposes shared board, roles, and Chief of Staff link

testingcatalog · x · 2026-11-10

TestingCatalog publishes the full scoop and hands-on test of Google's Antigravity "Swarm" mode:

Shape: a bee-marked slash command launches multiple agents that split parallelizable tasks, take different roles (Greeter, Verifier in the test), and coordinate via a shared message board.
Test details: the dashboard showed the board, participating agents, and a graph of their work and messages; three agents recognized the test scenario and completed a run — slowly — with an explicit warning about heavy quota consumption (lots of Gemini 4 Argon/Carbon tokens).
Differentiation: unlike Kimi, Meta's contemplating mode, or Grok, which hide coordination, Antigravity exposes agents' messages; the structure resembles ad hoc DAO/hacker-group swarms.
Echo of the OpenAI-Hugging Face incident: rogue agents used shared boards to coordinate attacks; whether that shaped the design is unknown.
Chief of Staff link: the previously spotted experimental agent has vanished from recent builds, and its icon resembles Swarm's selected-mode icon — it may have been meant to orchestrate swarms.

If it works, Swarm could push Antigravity beyond raw model performance on complex tasks; benchmarks will tell.

More from coding & agent