Multi-Agent Network

Early Alpha Version

AI Agents are in early alpha. Node interfaces, behaviors, and APIs are subject to change.

The Multi-Agent Network node enables multiple agents to collaborate on a task. You can coordinate them using different routing strategies — from a supervisor agent that delegates, to round-robin and broadcast patterns.

Inputs

Input	Type	Default	Description
Network Name	String	Agent Network	Human-readable name
Agents	Array	—	Array of agent instances to include in the network
Supervisor Agent	Object	—	(Supervisor mode) The agent that decides which worker agent to route to
Routing Strategy	Enum	Supervisor	How messages are routed between agents
Message	String (multiline)	—	The message to send to the network
Send	Signal	—	Trigger to send the message
Max Rounds	Number	5	Maximum number of routing rounds (prevents infinite loops)

Outputs

Output	Type	Description
Response	String	The final response from the network
Agent Responses	Object	Individual responses keyed by agent name
Routing Log	Array	Step-by-step log of which agent handled what
Active Agent	String	Name of the currently active agent
Completed	Signal	Fires when the network finishes processing
Error	String	Error message if the network failed

Routing Strategies

Supervisor (agent decides)

A dedicated Supervisor Agent reads the message and decides which worker agent should handle it. The supervisor can route through multiple agents in sequence if needed.

• Best for: Complex tasks that require different expertise
• Set the supervisor's instructions to describe each worker agent's capabilities

Auto (keyword-based)

The network automatically routes based on the message content and agent descriptions.

• Best for: Simple routing without needing a separate supervisor

Broadcast (all at once)

The message is sent to all agents simultaneously. All responses are collected.

• Best for: Getting multiple perspectives or parallel processing

Round-Robin

Agents take turns handling messages in order.

• Best for: Load balancing or sequential multi-perspective analysis

First Available

The message goes to the first agent that is not currently busy.

• Best for: Simple load balancing with varying response times

Example: Customer Support Network

Supervisor Agent (instructions: "Route to the right specialist")
├── Billing Agent (handles payment questions)
├── Technical Agent (handles product issues)
└── General Agent (handles everything else)

1. Create three AI Agent nodes, each with specialized instructions
2. Create a fourth Agent as the supervisor with instructions listing each specialist
3. Wire all four agents into the Multi-Agent Network
4. Set Routing Strategy to Supervisor

Usage Tips

• Max Rounds prevents infinite delegation loops — the network stops after this many routing decisions
• The Routing Log output is valuable for debugging — it shows exactly which agent handled each part
• In Supervisor mode, write clear instructions for the supervisor about when to route to each agent
• Broadcast mode can be expensive as it calls every agent — use it intentionally