Agentic Flows

Overview

An Agentic Flow is an Orchestrator application where AI agents — rather than usual deterministic building blocks — drive the interaction. Instead of following hard-coded paths through miniApps and condition blocks, agents use a large language model (LLM) to analyze caller intent and decide how to respond, which tool to invoke, or which specialized agent to delegate to.

This approach is suited for scenarios that require dynamic routing, multi-step task automation, or knowledge-base-driven question answering. Traditional deterministic flows can still be used when you need precise control over every conversation path.

Agentic Flows extend the capabilities of OCP-type applications. You can combine agents with deterministic flows and miniApps in the same application. A Concierge Agent can route callers to a Task Agent for payment processing, while a separate branch uses a traditional Flow for address collection.

An Agentic Flow typically follows this structure:

1. A Concierge Agent greets the caller and determines the intent.

2. The Concierge Agent delegates to one or more Task Agents (for workflow execution), QnA Agents (for knowledge retrieval), or Escalation Queues (for human handoff).

3. Each Task Agent uses tools (miniApps, MCP servers, web services) to complete its goal.

4. Each QnA Agent retrieves answers from a Pathfinder knowledge base.

This article documents the manual configuration workflow.

Agent Types

There are three agent types available for the Orchestrator applications. Each type has a distinct role within an Agentic Flow and enforces specific restrictions on the components it can connect to.

Concierge Agent

The Concierge Agent is the top-level routing agent and the first point of interaction for callers in the contact center. It analyzes caller intent and delegates to the appropriate sub-agent or escalation queue.

Key characteristics:

• Available only in Agentic-type applications.

• Uses other Agents, Pathfinder knowledge bases, Flows, and miniApps (except Intelli, WebService, and Announcement) as its Connected Tools.

• Manages the welcome prompt and escalation queues.

Task Agent

A Task Agent is a dedicated agent that executes a specific, often multi-step workflow with minimal human intervention. It uses tools — miniApps, MCP servers, and web service integrations — to achieve a goal.

Key characteristics:

• Also available in the OCP-type applications, where it can be added to canvas as a Flow, connecting to other miniapps and building blocks.

• Can only use Intelli, WebService, and Announcement types of miniApps and MCP servers as its Connected Tools.

• Connects as a sub-agent to the Concierge Agent.

QnA Agent

A QnA Agent answers caller questions by retrieving information from a Pathfinder knowledge base. The control plane manages the primary reasoning logic; Orchestrator provides the configuration for the Agent.

Key characteristics:

• Available in both Agentic-type and OCP-type applications.

• Connects to active knowledge bases from Pathfinder.

• Connects as a sub-agent to the Concierge Agent.

Read more on how to add a knowledge base to the agent below.

Agent Type Comparison

The following table summarizes the capabilities and restrictions for each agent type:

Creating Agentic Application

Agents operate within Agentic-type dialog applications. Create the application container before setting up any agents.

1. Navigate to the Orchestrator tab.

2. Click +Create to open the application creation dialog.

3. Provide the following details:

   • Application Name: enter a name for the new dialog application.

   • Group: select a group from the dropdown list. The dialog application is accessible to other group members.

   • App Type: select Agentic to enable agents for this application.

   • Description (optional): add a description. The maximum length is 255 characters.

   • Labels (optional): type a label name and press Enter to add it. Click the X icon to the right of a label to remove it.

4. Click Save.

The Orchestrator canvas opens. Agents appear as visual blocks on the canvas, similar to flows.

Creating Agents

After creating an Agentic application, set up the agents that compose the conversation logic.

1. Open the Agentic application.

2. In the building blocks menu on the left, locate the Agents section.

3. Click + Create.

4. On the Create Agent page, provide the following details:

   • Agent Name: enter a descriptive name.

   • Agent Type: select Concierge Agent, Task Agent, or QnA Agent.

   • Description: enter a summary of what the agent does.

   • Labels (optional): add one or more labels.

   • Input Fields: define the data parameters the agent receives (see Configuring Input and Output Fields below).

   • Output Fields: define the data parameters the agent returns.

5. Click Save. The agent setup panel opens, where you configure instructions, tools, escalation queues, and other settings.

Configuring Agents

The agent setup panel contains several configuration areas. The following sections describe each area in the order it appears in the interface.

Agent Instructions

Instructions are the natural-language directives that govern how the agent behaves, what tone it uses, and which rules it follows during a conversation.

To add instructions:

1. In the agent setup panel, locate the Instructions area.

2. Type an instruction and press Enter to save it as a separate line.

3. Add as many instruction lines as needed.

4. Double-click an instruction to edit it, or click the X icon to the right to delete it.

The instructions area is collapsible. Use the toggle to collapse or expand it when managing long instruction sets.

Input and Output Fields

Input and output fields allow agents to exchange information with each other and with the application canvas. An input field receives data from a preceding agent or the canvas. An output field passes data to the next agent or back to the canvas.

As for output fields, they consist of:

• Name: the identifier used for data mapping between agents.

• Output Field Type: the output data type (for example, string or number).

• Description: a dropdown text field including natural-language explanation of what the output field contains. The agent uses this description to decide how to interpret the field.

Input and output fields can be edited from within the agent configuration panel after the agent is created. Click the field to modify its name, type, or description.

Connecting Tools and Agents

The Connected Tools & Agents add to the capabilities the agent can invoke during a conversation. The available components depend on the agent type (see the Agent Type Comparison table above).

To connect a component:

1. In the agent setup panel, locate the Connected Tools & Agents area.

2. From the left-side menu, drag a component and drop it onto the area.

3. Double-click the added component to edit its Name and Description, and also add Input Fields and their corresponding descriptions.

4. To remove a component, click the trashcan icon next to it.

5. Click Save to preserve the configuration.
   

   

The tool description tells the agent when and how to use the tool. Write clear, action-oriented descriptions. For example: "Use this knowledge base to answer all general customer questions about hotel amenities and policies."

Knowledge Bases (Pathfinder)

Knowledge bases are available as components for Concierge Agents (as a connected tool) and QnA Agents (as the primary connected resource).

When connecting a Pathfinder knowledge base to a Concierge Agent, drag the knowledge base from the section in the left-side menu:

Click on the connected knowledge base to edit its Name, Description and add input fields (with descriptions):

Knowledge Base in QnA Agent

QnA Agent uses a Pathfinder knowledge base as its connected tool. When configuring the QnA agent, select the Pathfinder project from the dropdown list:

For the Pathfinder projects with the knowledge base enabled, additional filtering schema for the metadata can be set in Pathfinder. These filtering fields can then be used for the retrieval with the description provided so that the control plane understands how to apply this information.

For the Pathfinder projects without the enabled knowledge base, a vector store must be additionally selected from the list:

MCP Tools

MCP (Model Context Protocol) tools connect agents to external data sources and services through a remote MCP server.

To add an MCP tool:

1. In the left-side menu, open the MCP section under the Agents and Tools tab.

2. Click +Add.

3. Provide the MCP server URL and token (if applicable).

4. Click Add.

5. The MCP tool appears in the left-side menu. Drag it onto the Connected Tools & Agents area.

6. Click Save.

MCP tool metadata is not persisted locally. The list of MCP tools disappears from the left-side menu when you refresh the page. However, the tool connection in Connected Tools & Agents is preserved after you save.

Configuring Escalation Queues

Escalation queues define the conditions under which the Concierge Agent transfers the caller to a human operator. Each queue maps to a specific contact center destination, such as a RingCX queue.

To configure an escalation queue:

1. In the agent setup panel, locate the Escalate to Human section.

2. Click + Add Escalation Queue.

3. Provide the following details:

   • Queue Name: a display label for the queue (for example, "Technical Support").

   • Queue ID: the numeric identifier matching the contact center queue (for example, 791311).

   • Queue Description: a natural-language description of when to route to this queue. For example: "Redirect the caller here when they report a technical problem with their account."

4. Add additional queues as needed.

Escalation queues are configured on Concierge Agents. Task Agents and QnA Agents don't manage escalation directly — the Concierge Agent handles all human handoffs.

The Queue ID is used during testing to validate that the agent routes callers correctly. After a test conversation, check DiaManT logs for the "Extension value is:" entry to confirm the Queue ID matches the expected destination. See Test Dialog Application for details.

Setting Welcome Prompt

The welcome prompt is the first message the caller hears when the Concierge Agent starts a conversation. It sets the tone and identifies the AI assistant.

To configure the welcome prompt:

1. In the agent setup panel, locate the Welcome Prompt section.

2. Type the greeting message in the text field.

3. Click Save.

The welcome prompt is configured on Concierge Agents. Task Agents and QnA Agents don't display a welcome prompt — they receive control after the Concierge Agent has already greeted the caller.

Agent Settings

The agent setup panel includes a Settings area with configuration options that affect overall agent behavior. The following settings are specific to Agentic Flows.

Reset on Intent

When enabled, the Reset on Intent setting forces the agent to clear its accumulated conversational context when caller intent changes mid-conversation. This prevents stale context from one topic (for example, billing) from influencing the response to a new topic (for example, technical support).

To configure this setting:

1. In the agent setup panel, navigate to the Settings area.

2. Toggle Reset on Intent on or off.

For general information on intent-change behavior in Orchestrator, see Reset on Intent Change.

Configuring Event Handlers for Agents

Task Agents can be used in Event Handlers in the same way as regular Flows.

Specific Event Types (type of errors) that trigger the Error Handler are available for the Agents:

AgentEscalated

AgentOutputIncomplete

Read more regarding the event-handling configuration workflow and event types in Event Handlers.