Google Docs for AI Workflows: How We Built Real-Time Multiplayer Collaboration

Redis serves as the message broker and state store. When Socket.IO is scaled across multiple server instances (as it is in our Docker Compose setup), Redis Pub/Sub ensures that a change made on server A reaches users connected to server B. Redis also holds the shared canvas state, enabling:

• Late join — a user who opens the workflow after others have been editing sees the current state immediately
• Crash recovery — if a server instance goes down, the state is preserved and rehydrated
• Snapshot history — periodic snapshots enable time-travel debugging

What Gets Synchronized

Not everything on the canvas needs real-time sync. We made deliberate choices about what to broadcast:

Synced in real-time:

• Node positions (drag and drop)
• Edge connections (create, delete, reconnect)
• Node configuration changes (parameter updates)
• Block additions and removals
• Cursor positions (so you can see what teammates are working on)

Not synced:

• Zoom level and viewport position (this is personal)
• UI panel states (side panel open/closed, tab selection)
• Undo/redo history (each user has their own local history)

This balance keeps the protocol lean — we're sending tactical changes, not entire canvas snapshots on every keystroke.

Conflict Resolution

When two users edit the same thing at the same time, conflict is inevitable. Our approach:

Last-write-wins for simple values. If User A changes a node title to "Research" and User B changes it to "Analyze" a half-second later, the node becomes "Analyze." This is good enough for text fields because the visual feedback is immediate and users naturally coordinate.

Operational transforms for structural changes. Moving a node, creating an edge, deleting a block — these are operations applied in order. If User A deletes a node that User B is currently dragging, the operation queue handles it gracefully: the drag is cancelled, the node disappears from B's canvas.

Locking for critical sections. When a user clicks into a complex configuration form (e.g., the LLM provider settings panel), that form is "soft-locked" — other users see a "User X is editing this" indicator and can view but not modify until the lock is released.

Performance at Scale

Real-time sync is easy for two users on a 10-node workflow. It gets hard with 20 users on a 500-node workflow. We invested in:

• Delta compression — we send only what changed, not what stayed the same
• Batching — rapid successive changes (like nudging a node pixel by pixel) are batched into a single update every 50ms
• Binary protocol — Socket.IO messages use the binary transport for compact payloads
• Room management — users are only in rooms for workflows they have open. Closing a tab removes them from the room immediately

Why This Matters for AI Agents

Multiplayer isn't a "nice to have." For AI agent workflows in production, it's essential:

Review. Before an agent that handles customer data goes live, someone needs to review it. Multiplayer means the reviewer and builder can walk through the workflow together, in real-time, pointing at specific nodes.

Debugging. When an agent makes a wrong decision, the team gathers around the execution trace — just like developers gather around a stack trace. Everyone sees the same thing, at the same time.

Knowledge transfer. A senior engineer can "drive" while junior team members watch, ask questions, and learn. No screen sharing. No recorded videos. Just open the canvas together.

The Stack in Detail

The Socket.IO server runs as a custom Node.js server alongside Next.js — not embedded in the Next.js API routes. This gives us dedicated WebSocket handling without competing for Next.js request processing resources.

Redis is configured with AOF persistence for durability. Even if Redis restarts, active collaboration sessions survive.

What's Next

• Presence awareness — richer indicators of who's online, what they're doing, and what they've changed recently
• Comments and annotations — leaving notes directly on nodes and edges, like comments in Figma
• Change proposals — suggest a change to a workflow without modifying it directly; the owner reviews and accepts or rejects
• Video/voice integration — one-click call with everyone in the workflow room

Building AI agents is a team sport. The tools should be, too.

— Alexandr Chibilyaev