Google's Antigravity Launches Swarm to Run Parallel Coding Agents

Google has added /swarm to Antigravity, its agentic coding environment. The preview slash command launches named agents, assigns each one part of a task, and coordinates their work through a shared message board. It is currently available in the Antigravity App and CLI alongside Gemini 4 Carbon, Google’s long-horizon model.

Large changes can overwhelm a single coding agent as files, tests, and intermediate decisions accumulate in its context. Helper subagents add parallelism, but a primary agent still has to relay their findings and reconcile their output. /swarm distributes that coordination across named workers, verifiers, and a shared log.

One board, many worktrees

Entering /swarm <task> turns the current conversation into the swarm’s control surface. A coordinator divides the prompt into independent assignments and launches specialized workers in isolated Git worktrees. The separate checkouts reduce working-copy collisions, though integration can still produce merge conflicts when assignments overlap.

Google’s coordination model includes several distinct roles and safeguards:

• The coordinator names the swarm, divides work by file, package, or question, and creates workers such as @parser, @api-tests, and @docs.
• The coordinator typically creates one @verifier for every two to four workers. Verifiers inspect diffs, run tests, and report defects to the worker responsible for the code.
• Assignments, questions, diff summaries, and test results appear on an append-only board. Agents address @name, @group, or @all, keeping coordination in one ordered log.
• According to Google, idle agents park without consuming tokens and resume when another agent requests input.
• A silent @monitor intervenes when every member becomes inactive before the task has finished, prompting an agent that can unblock the work.

Assign models by role

The model picker can assign a different model to each agent. Google positions Gemini 4 Carbon for coordination, architecture, security audits, and adversarial verification because it is tuned for extended exploration and verification loops. Lighter models such as Gemini 4 Argon can handle routine edits with lower token use.

Google reports that Carbon ranks first on CursorBench 4.0 and Perplexity’s WANDR, benchmarks designed to test long-horizon coding agents and research systems. Those results support Google’s recommended use cases, though teams should validate model choices against their own repositories, tests, and quota limits.

Watch the agents coordinate

Antigravity exposes swarm activity in an App side pane and a full-screen CLI interface opened with /swarm board or Alt+S. The interface provides three views:

• Board: A live, append-only feed with filters and a New divider for unread posts.
• Members: A roster grouped into Active, Standing by, and Retired. Each row shows the model ID and latest activity, and developers can open a worker to steer it.
• Graph: An interactive timeline showing which agents mentioned one another. Arrow weights reflect post counts, and replay controls trace activity from the initial brief through the final diff.

Choose by task shape

Antigravity offers four multi-agent mechanisms for different task sizes and dependency patterns. /swarm fits work that can be divided into independent files, packages, or research questions over several minutes or hours.

Google says its engineers tested swarms for five weeks on multi-package migrations, cross-subsystem investigations, and architectural redesigns. Documentation examples include adding validation to three API handlers in parallel, assigning verification to a different model to check for race conditions, and grouping pytest failures by package before assigning one worker to each package.

Quota, recovery, and access

Parallel workers consume quota faster than a standard agent session, even when idle agents stop using tokens. If a member exhausts its quota, it pauses and does not resume automatically after the quota refreshes. Resuming it requires an addressed board post, a model change, or direct steering from the member view.

/swarm is available in the App and CLI on Google AI Pro and Google AI Ultra plans. Gemini 4 Carbon requires an Ultra subscription for individual users; Google lists Gemini for Business as the enterprise route.

Teams that currently assemble LangGraph or CrewAI pipelines for parallel code changes can use Antigravity’s built-in coordinator, worktrees, message board, and verifier roles instead. The feature remains in preview, so commands, interfaces, limits, and recovery behavior may change. Isolated worktrees also leave integration and code review responsibilities with the developer, especially when workers touch related code.