Type /swarm followed by a task in Google’s Antigravity coding tool, and your current conversation stops doing the work and starts managing it. According to Google’s Antigravity documentation, that conversation becomes a coordinator. It cuts the job into pieces, hires a named agent for each one, and lets them all run together.
The names are literal. A worker might be called @parser, @api-tests, or @docs, and similar workers can be numbered as a group, like @fix-1 up to @fix-4. Each starts with a blank memory in its own conversation, and its assignment is posted where everyone can read it: its territory, its off-limits areas, and the definition of finished. The usual split gives each file or package its own worker, so two agents never edit the same thing.
Checking is built in. Roughly one verifier goes in alongside every handful of workers (two to four, per the docs). It reads the changes, runs the tests, and returns problems to whoever authored the faulty code. Google’s examples include giving the checker a different model from the workers, so one model audits another’s output, and one asks it to run Go’s race detector. A different pair of eyes costs nothing extra to request. You just say so in the prompt.
Everything happens on a single shared message board. Assignments, questions, summaries of changes, and test results all land there, and workers never message each other privately. In the Antigravity App you watch it in a side pane with a live feed, a member list, and a graph of who addressed whom. A worker with nothing to do sleeps without spending tokens until someone mentions it, and a separate monitor worker speaks up only if the whole team goes quiet before the job is done. You can keep talking to the coordinator the entire time.
Google is specific about when to use it. The fit is work that divides cleanly: three API handlers in three files, failing tests grouped by package, or a hunt for why a command-line tool takes four seconds to start, with three read-only workers each chasing a different suspect. Its sibling commands cover other shapes. /boost is for problems that are hard rather than large, /teamwork is for projects that run hours to days across many milestones, and plain subagents handle one side task. The docs put a swarm in the range of minutes to hours.
The page lists the Antigravity App and the command-line version as the two places it runs. It does not say which subscription tier unlocks it, and it publishes no results on how often a swarm beats a single agent. The examples are all prompts, not measurements.
The documented warning is about quota. If a member runs out, it pauses, shows you the error once, and stays paused until a new post appears, you change its model, or you write to it directly. It does not pick up again when your quota refreshes. Because every worker is its own conversation on a model, a swarm of four workers and a verifier means five conversations drawing on one allowance, which is a very different burn rate from a single chat.
Start with a job you could split by hand in two minutes, like the multi-package test repair in Google’s examples, and watch how fast your quota moves before you hand a swarm anything bigger.
Reported by Google Antigravity documentation (undated page on the /swarm command), as read on 11 November 2026.
