Skip to main content

Route Conversations with a Condition Agent

Real bots rarely do just one thing. A support bot needs to tell a billing question from a technical question and handle each differently. In this tutorial you'll use a Condition Agent to read the visitor's message and send the conversation down the right branch — including the one branch beginners always forget: an explicit catch-all for messages that match nothing.

What you'll build: a triage bot that routes messages into Billing, Technical, or Everything else.

Nodes you'll use:

Concepts to know:

The one rule you must not skip​

A Condition Agent creates exactly one output branch per scenario you define — and there is NO automatic Else. If a message matches none of your scenarios, and you didn't add a catch-all scenario, that message is silently dropped. The visitor gets nothing back.

Always add an explicit default

Every Condition Agent needs a final "other / anything else" scenario. This is the single most common mistake in AgentFlow. If in doubt, add the catch-all first, then add your specific scenarios above it.

Step 1 — Start the flow​

Create a new AgentFlow. Open the Start node and set Input Type to Chat Input. Connect the Start node to a new Condition Agent node.

Start connected to a Condition Agent that routes by intent

Step 2 — Configure the Condition Agent​

Open the Condition Agent and set three things:

  1. Model — pick an AI model (see Models).

  2. Input — what the agent should evaluate. Set it to the visitor's message:

    {{$question}}
    The $ is required

    It must be {{$question}}. The dollar-less form {{question}} never resolves, and the agent will have nothing to classify.

  3. Instructions — a short description of the routing job, for example:

    Classify the user's message into one of the scenarios below based on their intent.

Condition Agent Input field reading the visitor's question

Step 3 — Define the scenarios (branches)​

Under Scenarios, add one entry per route. Each scenario becomes its own output branch on the node. Add them in this order:

#ScenarioMeaning
1The user has a billing, payment, invoice, or refund question→ Billing branch
2The user has a technical problem, bug, or how-to question→ Technical branch
3Anything else, or the intent is unclear→ Catch-all branch

Scenario 3 is your explicit default — it guarantees no message is ever dropped.

Condition Agent scenarios with an explicit catch-all branch

One output per scenario

Notice the node now has three output handles — one for each scenario, in order. There's no fourth "else" handle. The catch-all you wrote is the else.

Step 4 — Build a branch for each route​

For each output handle, add an Agent → Direct Reply pair, and connect the matching Condition Agent output to it.

  • Billing branch: an Agent with a System message like "You handle billing questions for Acme. Be precise about invoices, refunds, and payment methods."
  • Technical branch: an Agent whose System message focuses on troubleshooting.
  • Catch-all branch: an Agent that gives a friendly general reply and offers to connect the visitor to a human.

Name each Direct Reply after the route it serves — the finished flow below shows Complaint and Anything else — so the canvas stays readable once the branches fan out.

Each Direct Reply carries its own Agent's answer by referencing that Agent's label, not a generic output. Type {{ in the Message field and pick the Agent sitting on the same branch; the labels are per-node, so a branch pointing at the wrong Agent sends the wrong answer.

Turn Show Output in Chat off on every Agent, otherwise each answer is sent twice — once by the Agent, once by its Direct Reply. See Show Output in Chat.

Skip the Agent for canned routes

A branch whose answer never changes doesn't need an Agent at all — connect the Condition Agent output straight to a Direct Reply holding the fixed text.

Your finished graph fans out from the Condition Agent into three parallel paths.

The full conditional routing flow from Start to the branches

Step 5 — Test each path​

Open the chat tester and send one message per route:

  • "My invoice looks wrong." → Billing branch.
  • "The app crashes on login." → Technical branch.
  • "Do you ship to Canada?" → Catch-all branch.

Confirm each reply comes from the right Agent. Then send something deliberately weird ("banana") — with the catch-all in place, it still gets a graceful answer instead of silence.

What you learned​

  • A Condition Agent reads a message and routes it, creating one branch per scenario.
  • There is no implicit Else — you must add an explicit catch-all scenario or unmatched messages are dropped.
  • The classification input must use {{$question}} with the $ prefix.
  • Each branch is its own mini-flow (Agent → Direct Reply) tailored to that intent.

Next steps​