Google Antigravity adds /boost: a multi-agent mode for hard tasks
It fans a prompt out to subagents that plan, implement, and test in parallel, then loops on failures until checks pass — paid plans only, in the app and CLI.
Copy markdown
What /boost actually runs
A Primary Orchestrator reads your prompt and workspace, then splits the job into discrete, verifiable subtasks. Specialized subagents implement, investigate, and run local tests in isolated scopes in parallel; the orchestrator then synthesizes results, runs regression checks, and loops on the diagnostics until every test passes.
When it's worth the extra tokens
Google aims it squarely at the tasks single-pass agents flail on: concurrency bugs and race conditions, algorithmic problems, complex multi-file refactors, and root-cause hunts across unfamiliar code. Task horizon runs 'seconds to hours,' so it's overkill — and pricier — for quick one-file edits.
The catch: paid-only, and familiar
Boost is limited to paid Antigravity 2.0 and the Antigravity CLI — no free tier — and spinning up multiple agents burns extra quota. Early Hacker News readers note the pipeline mirrors Claude Code-style subagents and grumble at putting deeper reasoning behind the paywall.