Google announced a preview of /swarm on November 10, 2026, for developers using the Antigravity App and command-line tool. It divides a coding task among AI workers that coordinate on a shared message board, with another agent checking their changes.
For software teams, the useful question is whether parts of a job can move forward independently. Our recommendation: start with one agent for a contained edit; consider a swarm when you can assign distinct responsibilities and check the combined result.
The app changelog already lists /swarm under November 9 and says releases may take a few days to reach all users. The announcement is a preview rollout, not confirmation that every account has it.
Choose by dependencies, not file count
The Swarm guide recommends work that splits across independent files, packages or questions. Apply that distinction to the job in front of you:
One small edit, or a chain of dependent decisions: start with one agent. If each step needs the previous result, there is little work to do in parallel.
Several bounded changes: consider a swarm when ownership is clear. An illustrative assignment would give separate workers unrelated test gaps in different modules, then require a check of the combined changes.
An investigation with several plausible causes: consider read-only workers. For example, investigate a slow internal dashboard’s data retrieval and page rendering separately, asking each worker for supporting evidence before choosing a fix.
File count alone is a poor guide. A ten-file change to one shared interface may need more coordination than several unrelated fixes. Agree on that interface first, then identify what can proceed independently.
Swarm is not Antigravity’s first multi-agent option. It already offers background subagents. Google positions /boost for difficult bugs and reasoning-heavy work, while /teamwork adds an upfront scoping interview and milestone checks for larger projects. Choose the coordination style that fits the problem, not simply the largest team.
Check where each worker will edit
Workers share a workspace by default; separate working copies are optional. That is the rule in the detailed Swarm guide, so do not assume that launching a team automatically separates its edits.
Before parallel editing, establish who owns each file and shared interface. Separate copies can keep simultaneous edits apart, but the changes still need to work together. Ask for the combined result to be checked before accepting it.
There is also a cleanup consequence: removing a worker deletes its separate working copy. Have its changes integrated before removing it. A worker finishing its assignment is not, by itself, proof that those changes are safely in the final project.
Budget for the whole job, including a pause
Google warns that parallel teams consume quota faster than ordinary chats. Its subagent documentation says idle board waits consume no tokens. Free waiting does not establish that a completed team task will be cheaper or faster; the launch provides no controlled comparison with a single agent.
The plans page says usage depends on the work performed; remaining model quotas can be viewed in Settings. Pro and Ultra users can spend purchased AI credits after the baseline allowance is exhausted. The AI Credit Overages setting offers Never or Always; Always allows automatic credit use. Check that setting before starting, rather than assuming your included allowance is a spending cap.
A quota-paused worker does not restart automatically when quota refreshes. The documented wake-up events are a new board post, a model switch or a direct message. Check remaining capacity before resuming: waking the worker does not create more quota, and another model may have its own limits.
Keep model access separate from team access. Google’s model matrix lists Gemini 4 Carbon for Ultra and Enterprise, not Free/Plus or Pro. The Swarm guide says workers inherit the conversation’s model and can use alternatives available in your model picker. Those facts do not establish that every swarm requires Carbon.
The documentation does not spell out a swarm-specific minimum plan or complete regional availability. Confirm access in your account before upgrading for this feature. There is no fixed price per completed swarm in these sources.
Decide what counts as finished before starting
For a first assignment, specify the outcome, which files each worker may change, what must remain untouched and the evidence needed for acceptance. For example, if workers add tests to separate modules, require the combined test suite to run without weakening existing assertions.
Use the board and worker conversations to inspect progress. Before accepting the result, have a person review the combined code changes and actual test output against the original request. An AI verifier’s summary is evidence to examine, not permission to release.
The same distinction between seeing agent activity and approving code is explored in RohitAI’s guide to Copilot’s review canvas in Teams. For Antigravity, start with a bounded job whose success you can check; expand the team only when the work has genuinely independent parts.