Triage agent
Learn what the triage (master) agent does, how it decides where each chat or ticket goes, and how to control its routing with your own instructions.
The triage agent is the first AI agent to read a new chat or ticket. It works out what the customer needs and sends the conversation to the agent best placed to handle it: one of your sub-agents, the Knowledge Base agent, or a human. It usually does not answer the customer itself.
In the dashboard the triage agent is called the master agent. Every workspace has two, and Robylon creates them for you: one for chat, which can work on all your chat channels as required, and one for email on the Robylon Email channel. You do not create master agents yourself. In the AI Agents list each one appears as Multi Agent with a Master Agent tag, often simply named “Triage Agent”.
Open it from Settings → Workflows & AI Agents → AI Agents.
How the triage agent routes conversations
It checks your instructions first, then matches the message against the agents it can reach, then falls back to the Knowledge Base agent.
- Read the conversation. The agent reads the latest message together with the conversation so far and any customer details in User Info.
- Apply your rules. Rules in its instructions come first. Use them for checks that must happen before anything else, for example spotting spam.
- Match an agent. The triage agent gets one handover tool for each agent it can reach, built from that agent’s purpose, when-to-use criteria and example phrases. It picks the best match and calls that handover tool. See how sub-agents are described to the triage agent.
- Fall back. If nothing matches, it hands over to the Knowledge Base agent, unless your instructions name a different fallback. It asks the customer to clarify only when it cannot tell what they want at all.
It takes one action per turn, so each message is routed to exactly one place.
flowchart TD A[New chat or ticket] --> B[Apply your routing rules] B -->|Rule matches| C[Do what the rule says: route, hand to a human, or close] B -->|No rule| D[Match against each agent's description] D -->|Match found| E[Hand over to that sub-agent] D -->|No match| F[Hand over to the Knowledge Base agent] E -->|Out of scope| B
Agents it can route to
By default, every live sub-agent on the same channel, plus the Knowledge Base agent. Open the triage agent’s Agents section to see the list and limit it if some agents should not be reachable from this channel.
Handing over to a human is done through a tool, for example a human agent handover tool, added under Tools Available. The triage agent only tells a customer they are being transferred when it actually calls such a tool.
Set up the triage agent
Build the agents it routes to, then tell it where each kind of request should go.
Build your sub-agents first
Create and publish a sub-agent for each area you want handled, for example returns. See sub-agents.
Open the triage agent
In Settings → Workflows & AI Agents → AI Agents, open the agent tagged Master Agent for chat or for email.
Choose the agents it can reach
Open the Agents section. By default it can route to every live sub-agent on its channels. Limit the list if some sub-agents should not be reachable.
Write the routing instructions
Click Edit, and in Custom Instructions describe which requests go to which sub-agent, how to break ties, and what to do when nothing matches. See how to write routing instructions. Then publish the new version.
One triage agent per channel
There is one master agent for chat and one for email. The chat master agent can work on all your chat channels as required, and the email master agent works on the Robylon Email channel.
Triage agent settings
| Section | What it does |
|---|---|
| Custom Instructions | The routing rules: which requests go where, tie-breaks, what to handle directly and the fallback. |
| Agents | The sub-agents it can hand conversations to, with a count of how many are selected. |
| Tools Available | Tools it can call itself, for example ending a chat. |
| User Info | Customer variables it sees on every turn, for example the customer’s email. |
| LLM Configuration | The model, temperature, max tokens and thinking tokens. See custom agents. |
Write good routing instructions
Tell the triage agent exactly how to sort requests, in the order it should check them. Good routing instructions cover five things:
- The role. “You route each message to the right agent. You do not answer questions yourself, except the cases listed below.”
- Checks in priority order. For example, first close spam, then hand over anyone asking for a person, then route by topic.
- A routing map. For each sub-agent, the kinds of request it owns, with a few example phrasings.
- Tie-breaks. Rules for messages that could fit two agents. For example: “If the customer mentions a refund, always route to the returns agent, even if they also mention delivery.”
- The fallback. What to do when nothing fits, for example routing to the Knowledge Base agent.
An example for an online store:
Route to the Returns agent for returns, exchanges, refunds, and damaged or wrong items. Route to the Orders agent for order tracking, delivery delays, cancellations before dispatch, and address changes. If a message mentions both a refund and a delivery, route to Returns. For general questions about products or policies, route to the Knowledge Base agent. If the customer asks for a person, hand over to a human.
When a sub-agent cannot help
Every sub-agent has a built-in way to hand the conversation back to the triage agent, for example when the customer changes topic. The triage agent then routes the new request again. See sub-agents.