Google Antigravity adds a /boost multi-agent reasoning mode

Sources: primary, discussion
Summary: The documentation describes a three-phase pipeline. An orchestrator forms a strategy and splits the task, dispatches implementation and investigation subtasks to subagents running in ephemeral isolated worktrees, then validates the combined result against full test suites and feeds failed assertions back for another round. Boost is documented as sitting between the default single-agent loop and the multi-day Teamwork mode, with a stated horizon of seconds to hours, and it is limited to paid plans on Antigravity 2.0 and the CLI. Subagents inherit the workspace file access and command permission policies rather than bypassing them.
Why it matters: Subagents inheriting the parent's file access and command permissions is the detail that decides whether a fan-out mode widens the blast radius of an agent run.