Antigravity Swarms
12-Factor Analysis Decision Framework

Deep Comparative Analysis of All 5 Solutions

An exhaustive, side-by-side engineering evaluation across latency, throughput, reactive wake-up, zero-trust security, and operational complexity.

1. Comprehensive 12-Dimension Engineering Matrix

Dimension 1. Native Remote Control 2. Programmatic AgentAPI 3. Networked MCP Bus 4. Tailscale + NATS 5. Fleet Cockpit (AgentsRoom)
Transport Protocol HTTPS :443 WebChannel (XMLHTTP Stream) SSH / Local UNIX Socket / HTTP Sidecar Streamable HTTP / SSE / WebSocket WireGuard UDP + NATS Binary Protocol E2EE WebSocket Relay (:443)
Latency (LAN / WAN) 100–300ms (Cloud round-trip) <5ms LAN / 20–60ms WAN 10–30ms LAN / 30–80ms WAN <1ms LAN / 15–40ms WAN 50–100ms (Relay round-trip)
Throughput / Capacity Low (single streaming channel) Medium (10–50 turns/sec) High (Redis in-memory speed) Extreme (100k+ msg/sec) Medium (human-paced stream)
Reactive Wakeup Semi-auto (Browser / Cloud UI) Full (`agentapi send-message`) Poll-based (`poll_inbox`) or Hybrid Full (Consumer Group Push) Webhook / CI / Mobile Trigger
NAT / Firewall Traversal Google Cloud Relay Requires open SSH or VPN Standard HTTP Reverse Proxy 100% Auto (STUN/DERP) Built-in E2EE Relay
Setup Complexity Zero (Built into agy) Low (CLI commands) Moderate (FastMCP + Redis) Moderate (Tailscale + NATS) Low (SaaS / CLI binary)
Message Durability Ephemeral stream Stored in trajectory logs Durable in Redis/Postgres Durable (NATS JetStream DLQ) Durable in Cockpit Cloud DB
Conflict Prevention None (manual user oversight) Manual script coordination Atomic Distributed Locks Subject partitioning & KV locks Automatic Git Worktrees
Security & Zero-Trust Google OAuth Identity SSH Keys / Bearer Token Bearer Token + SSL WireGuard Crypto Identity End-to-End Encryption (E2EE)
Human Observability Official Web UI (antigravity.google) CLI / Terminal outputs MCP Inspector / Redis CLI NATS CLI / Custom Dashboard Mobile & Web Cockpit + Diffs
Agent Heterogeneity Antigravity only Antigravity only Universal (Antigravity, Claude, Cursor) Universal (Any NATS client) Multi-Agent (Antigravity, Claude, Aider)
Operating Cost Free (Included in Antigravity) Free (Open-source tools) Free self-hosted / $5 VPS Free Tailscale & NATS tier Free / Paid SaaS Tier

2. Architectural Decision Tree: Which Solution Should You Choose?

graph TD Start["What is your primary goal?"] --> Q1{"Do you need human visual
monitoring & mobile control?"} Q1 -- Yes --> Cockpit["Choose Method 5: Fleet Cockpit (AgentsRoom) / Antigravity Web UI
• Mobile steering & voice dictation
• Git worktree isolation
• Webhook triggers"] Q1 -- No --> Q2{"Are you orchestrating 2-3 machines
or a 10+ node enterprise swarm?"} Q2 -- "2-3 Machines (e.g. MacBook + dell5040)" --> Q3{"Do agents need shared tools & file locks
or just direct task triggering?"} Q3 -- "Direct Triggering" --> AgentAPI["Choose Method 2: Programmatic AgentAPI & SDK
• Deterministic `agentapi send-message`
• Reactive wakeups via SSH
• Zero extra daemons"] Q3 -- "Shared Tools & Locks" --> MCP["Choose Method 3: Networked MCP Bus
• FastMCP over SSE (:8080)
• Redis mailboxes & Redlock
• Cross-tool interoperability"] Q2 -- "10+ Node Production Swarm" --> NATS["Choose Method 4: Tailscale Mesh + NATS JetStream
• Sub-ms WireGuard routing
• Guaranteed at-least-once task queues
• Automatic NAT hole punching"]

3. Real-World Deployment Scenarios

Scenario A: The Dual-Machine Developer (MacBook + dell5040 Home Server)

Recommendation: Method 2 (AgentAPI via SSH/Tailscale) + Method 1 (Remote Control Web UI)

Install Tailscale on both machines. Keep agy-remote-control.service running on dell5040 so you can view trajectories at antigravity.google. When your laptop agent needs to offload heavy compilation or server audits, it executes ssh dell5040 "agy agentapi new-conversation '...'". Zero extra servers to maintain.

Scenario B: Heterogeneous Agent Team (Antigravity + Claude Code + Cursor)

Recommendation: Method 3 (Networked MCP Bus with Redis)

Because Antigravity, Claude Code, and Cursor all support the Model Context Protocol, you can run a single FastMCP server that exposes send_agent_message, check_inbox, and acquire_distributed_lock. The agents pass tasks across different LLM vendors seamlessly without proprietary lock-in.

Scenario C: High-Volume Autonomous Swarm (CI/CD, Nightly Audits, 10+ VMs)

Recommendation: Method 4 (Tailscale + NATS JetStream)

Direct HTTP/SSH calls will suffer timeouts when 10+ agents execute concurrently. NATS JetStream decouples task assignment, providing at-least-once message durability, dead-letter queues, and automatic load balancing across worker consumer groups.

Scenario D: The Mobile Tech Lead Overseeing Autonomous PRs

Recommendation: Method 5 (AgentsRoom Cockpit + Git Worktrees)

Inbound GitHub webhooks wake headless daemons into dedicated Git worktrees. You inspect diffs and approve critical merges directly from the AgentsRoom iOS/Android companion app while away from your desk.

← Method 5 (Fleet Cockpits) Back to Main Overview →