Aller au contenu

Multi-Agent & Orchestrator

Ce contenu n’est pas encore disponible dans votre langue.

The Multi-Agent feature lets you create a master orchestrator that automatically routes conversations to the right specialized agent based on what the user is asking about.

ScenarioSetup
One product, general supportSingle agent (no orchestrator needed)
Multiple departments (sales, support, billing)Orchestrator + 1 agent per department
Multiple brands under one accountOrchestrator + 1 agent per brand
Complex product linesOrchestrator + 1 agent per product line
User message
Orchestrator Agent (routing)
↓ determines best match
Specialized Agent (answers)
Response to user
  1. The user sends a message to a single entry point (one WhatsApp number, one chat widget).
  2. The orchestrator agent analyzes the message and decides which specialized agent should handle it.
  3. The message is routed to the selected agent, which has its own knowledge base, triggers, and personality.
  4. The specialized agent generates the response.

The user never knows there are multiple agents — it feels like one seamless conversation.

Create one agent for each area of expertise:

  • Sales Agent — knows about products, pricing, promotions
  • Support Agent — knows about troubleshooting, returns, technical issues
  • Billing Agent — knows about invoices, payments, subscription changes

Each agent should have its own:

  • Knowledge base (only documents relevant to its domain)
  • System prompt (personality and scope boundaries)
  • Triggers (actions specific to its domain)

Create a new agent that will act as the router:

  1. Go to Agent Settings on the orchestrator.
  2. Enable the Orchestrator toggle.
  3. Configure which agents are available for routing.

In the orchestrator’s settings, define how messages get routed:

FieldDescription
Available agentsWhich specialized agents the orchestrator can route to
Routing promptInstructions for the AI on how to decide (e.g., “Route billing questions to billing-agent, technical issues to support-agent”)
Default agentFallback when no clear match is found

Point your channels (WhatsApp, Live Chat, Instagram) to the orchestrator agent, not the specialized agents directly. The orchestrator handles the routing transparently.

The orchestrator uses AI to understand intent and route accordingly:

  • First message: Routes based on initial intent.
  • Follow-up messages: Stays with the same agent unless the topic clearly changes.
  • Ambiguous messages: Routes to the default agent or asks a clarifying question.
  • Keep orchestrator knowledge minimal — It only needs to understand what each agent handles, not the actual answers.
  • Write clear routing prompts — List each agent and when to route to it. Be specific about boundaries.
  • Set a good default — Choose the most general-purpose agent as fallback.
  • Don’t overlap knowledge bases — Each specialized agent should own its domain exclusively.
  • Test with edge cases — Try messages that could go to multiple agents and verify routing.
Orchestrator prompt:
"You are a routing assistant. Route messages as follows:
- Questions about products, sizes, availability → products-agent
- Questions about orders, tracking, returns → orders-agent
- Questions about payments, invoices, refunds → billing-agent
- General greetings or unclear intent → products-agent (default)"
User messageRoutes to
”Do you have this shirt in blue?“products-agent
”Where is my order?“orders-agent
”I need a refund”billing-agent
”Hello!“products-agent (default)

The orchestrator’s routing decisions are logged. You can review which agent handled each conversation in the Conversations page. This helps you:

  • Identify misroutes and improve routing prompts
  • Understand traffic distribution across agents
  • Spot knowledge gaps in specialized agents