Route between specialist Agents¶
Use an Agentic Flow to direct requests to specialist Agents. For example, one Agent can help prospective buyers while another supports existing customers. Each Agent uses its own prompt and tools.
1. Define destinations¶
Create focused Agents and map each intent key to its destination. Include DEFAULT for unclear or unsupported requests. For example:
| Intent | Meaning | Agent task |
|---|---|---|
SALES |
New purchase or service inquiry | Explain approved options and collect interest. |
SUPPORT |
Help with an existing service | Gather the issue and follow the support process. |
DEFAULT |
Unclear or unsupported request | Ask a clarifying question. |
Use exact intent keys consistently. The DEFAULT Agent also supplies the Flow's opening greeting.
2. Set a shared Persona¶
Enter a Persona in the Flow settings if all participating Agents should share an identity and tone. Include {persona} in each Agent's prompt.
A filled Flow Persona replaces the individual Agents' Personas. Leave the Flow field empty if each Agent should use its own Persona. Keep each Agent's task in its system prompt.
3. Choose the routing mode¶
| Mode | Configuration |
|---|---|
| LLM | Routing model and prompt that selects a configured intent |
| Decision | Routing instructions, an explanation of every intent and its own conversation-history limit |
For Decision, add DEFAULT and at least one other intent. Give each a destination Agent and a description. Set the history limit between 0 and 100. If Decision is missing from the selector, it is not enabled for your workspace.
Before switching modes, review the instructions and descriptions for the new mode. The assigned Agents stay the same.
4. Handle follow-ups and topic changes¶
Choose SALES for a new purchase inquiry and SUPPORT for an existing-service issue.
For unclear requests, choose DEFAULT so the reception Agent can clarify.
Keep the previous intent for short answers to the current Agent's question.
A clear new caller intent takes priority over the previous intent.
Return only an intent key present in the configured mappings.
LLM routing prompts can use {conversation_history}, {conversation_fact_summary} and {previous_intent}. The previous intent is empty on the first turn. For Decision, describe each request type in its intent field.
5. Connect and test¶
Route a phone number to the Flow. Assign processing profiles to the Flow entry point where needed. Test every intent, an ambiguous opening, a short follow-up and a topic change.
If a test selects the wrong Agent, adjust the routing rule or intent description. If the correct Agent gives a wrong answer, update its prompt. Use voice debugging to see which Agent was selected.