Antigravity Swarms
Research Verified 7-Page Engineering Reference

Google Antigravity Multi-Instance Communication

A comprehensive technical architectural investigation detailing how multiple Google Antigravity instances (equipped with remote control daemons / agy --remote-control) can discover, talk to, coordinate tasks with, and steer each other across machines.

Solutions Evaluated
5 Patterns
Each with a dedicated technical page
Native WebRTC Status
Flag-Gated
Active: Outbound HTTPS WebChannel Relay
IPC Latency Envelope
<1ms – 250ms
Direct WireGuard vs. Cloud relay
Autonomous Wakeup
Full Support
Reactive turns via `agy agentapi`

Executive Architectural Takeaway

Binary disassembly of ~/.local/bin/agy confirms that Google built a complete WebRTC peer-to-peer data channel mesh into the Antigravity daemon (MeshDaemon, InitiateMeshSession, ProxyCommand). However, in current production builds, server-side Mendel configuration flags disable direct P2P mesh connections (WebRTC P2P mesh disabled by flag; using WebChannel transport only), forcing cloud traffic through an outbound HTTPS relay at jetski-webchannel.googleapis.com:443.

To achieve high-performance, autonomous, and reliable multi-agent swarms across your fleet (e.g. dell5040 server + laptops + cloud VMs), explore the five dedicated method pages below.

Explore the 5 Solutions (Dedicated Pages)

Method 1 Flag-Gated

Native Cloud Remote Control & Mesh

Deep analysis of agy-daemon, server.go, and WebChannel streaming. Explains how Antigravity connects to Google's cloud plane, why P2P WebRTC is currently flag-gated off, and how instances proxy commands.

View Method 1 Details →
Method 2 Deterministic

Programmatic AgentAPI, SDK & Sidecars

The official CLI/SDK IPC pipeline. Details agy agentapi send-message for reactive wake-up, new-conversation for isolated tasks, Python SDK harnesses, and persistent sidecar listeners.

View Method 2 Details →
Method 3 Tool Sharing

Networked Model Context Protocol (MCP) Bus

Connecting multiple Antigravity instances to a shared FastMCP server over SSE/HTTP backed by Redis. Provides mailboxes, cluster presence, and distributed CAS/Redlock locks to prevent file collisions.

View Method 3 Details →
Method 4 Production Swarm

Tailscale WireGuard Mesh + NATS JetStream

The enterprise-grade standard. Combines zero-trust WireGuard mesh networking (sub-millisecond wire speed, automatic NAT traversal) with NATS JetStream durable consumer groups and at-least-once task delivery.

View Method 4 Details →
Method 5 Human Cockpit

Multi-Agent Fleet Cockpits (AgentsRoom)

Local-first fleet management via AgentsRoom and Antigravity Web UI. Implements isolated Git worktrees to allow 10+ concurrent agents to edit code without Git lock collisions, with mobile steering and webhooks.

View Method 5 Details →
Comparison 12-Factor Matrix

Deep Comparative Analysis & Decision Tree

Direct side-by-side scorecard comparing all 5 solutions across latency, throughput, setup overhead, zero-trust security, and failure modes, complete with an interactive architectural decision tree.

View Full Comparison →

Quick Architecture Reference Table

Method Primary Transport Latency Reactive Wakeup NAT / Firewall Best Fit
1. Native Remote Control HTTPS :443 WebChannel (Relay) 100–300ms Semi-Auto Google Relay Web browser steering of headless daemons
2. Programmatic AgentAPI SSH / Local UNIX Socket / HTTP <5ms LAN Full (`send-message`) Requires SSH / VPN Deterministic script/pipeline automation
3. Networked MCP Bus Streamable HTTP / SSE / WebSocket 10–30ms Poll / Hybrid Standard HTTP / Proxy Heterogeneous agents sharing mailboxes & locks
4. Tailscale + NATS WireGuard UDP + NATS JetStream <1ms LAN Full (Consumer Group) 100% Auto (STUN/DERP) Production swarms requiring durable queues
5. Fleet Cockpits Local-first E2EE WebSocket Relay 50–100ms Webhooks / CI Built-in E2EE Relay Dev teams needing mobile UI & Git worktrees