Multichannel agents

Build agents that work across voice and webchat from a single project.

A multichannel agent is a single virtual assistant that works consistently across different channels — such as voice and web chat — while adapting to each channel’s unique needs. The design separates what stays the same (core behavior) from what changes depending on the channel (greetings, formatting, available actions).

How voice and chat differ

Voice and web chat interactions differ in meaningful ways:

• Voice is shaped by latency and audio limitations — users speak casually, may interrupt, and prefer short responses
• Web chat is more deliberate — users type at their own pace, can scroll back, and expect clear visual structure. Chat also enables rich UI elements like buttons and quick replies that aren’t possible in voice

Despite these differences, the core agent logic remains the same across channels.

Core principles

Keep core behavior unified

The core behavior of your agent should remain consistent across all channels. This includes:

• Prompting logic and knowledge base
• Reasoning steps and flows
• Metrics and reporting

Handle channel differences in one place

Channel-specific differences — such as greetings or formatting — should be handled in a single location. This avoids duplicating code and makes updates easier.

The platform automatically selects capabilities based on the channel. For example, voice-specific delays are turned off automatically for web chat at runtime.

Channel-specific greetings

Each channel can have its own greeting message. Store greetings as variables so they’re easy to update.

• Web chat greetings might include emoji and HTML formatting
• Voice greetings should use a more conversational tone

Set channel variables in your start function:

def start(conv):
    if conv.channel_type.startswith("webchat"):
        conv.state.greet_message = (
            f"Hi, you're speaking with Poly, a virtual agent for "
            f"{conv.variant.webchat_site_name} 👋<br><br>How may I help you today?"
        )
    else:
        conv.state.greet_message = (
            f"Hi there, you're speaking with a virtual assistant for "
            f"{conv.variant.site_name}. How can I help?"
        )

LLM and style prompt

You don’t need to manually add style guidelines to your behavior prompt for voice. A prompt decorator is automatically applied per channel — for example, the Raven voice model includes built-in style guidelines.

Configure which LLM powers each channel in Agent > Model use or via the experimental configuration.

Knowledge base and links

• To share a URL in chat, include it directly in the knowledge base topic — the system presents it correctly inline
• SMS is only available for voice, so don’t rely on it for other channels
• Keep your content as channel-agnostic as possible to minimize maintenance

Where voice sends an SMS with a link, webchat can display the link or content inline instead. Adjust the action per channel when needed.

Metrics and reporting

The same metrics system applies across all channels. Metrics designed for voice (like transfer rates or conversation length) also apply to web chat. All results are stored in the same database for unified reporting.

Deploy and test

1. Configure chat settings

   From Channels > Chat, configure your greeting, LLM, and style prompt for the chat channel.

2. Deploy the widget

   From Channels > Chat > Widget, generate a script tag and embed it in your website’s HTML headers. See the widget deployment guide for details.

3. Test with preview

   Use the Preview button to test your webchat agent. Verify that formatting displays correctly and channel-specific behaviors work as expected.

Summary

Concern	Approach
Core agent behavior	Unified across channels
Greetings and formatting	Channel-specific variables
LLM model	Raven for voice, GPT for webchat
Style guidelines	Handled by prompt decorator per channel
Metrics	Same system, unified database
SMS actions	Voice only — use inline content for chat