Agent

Define your agent’s personality and role.

Use the Agent page to control your agent’s character – its personality and role. These two fields shape how the LLM phrases responses across every conversation.

Configure personality and role in Build > Agent. The same page also contains the Behavior section for hard rules and constraints.

Where is the greeting? The greeting (the agent’s opening line) is configured per-channel under Channels > Voice > Voice configuration for voice and Channels > Chat > Chat configuration for webchat. On legacy projects without channel-specific settings enabled, the greeting still appears as a section on the Agent page.

Greeting

The agent’s opening line is configured per-channel:

• Voice greeting – set under Channels > Voice > Voice configuration.
• Webchat greeting – set under Channels > Chat > Chat configuration.

The greeting goes directly to TTS without LLM processing. Write it exactly as you want it spoken – the LLM cannot rewrite or adjust it at runtime. See conversation flow for why this matters.

You can include tool calls and variant attributes in the greeting to make it dynamic – for example, to greet callers by location or time of day.

If you need to override the greeting at runtime based on caller data (e.g. personalized “Welcome back, [name]” messages), return an utterance from your start function instead.

Personality

This field sets the tone and communication style across every response. Pick one or more built-in tags – Polite, Kind, Funny, Energetic, Calm, Thoughtful – or select Other to write a free-form personality string that matches your brand voice.

How the tags are used

The selected adjectives are joined together and inserted into the system prompt as a single sentence of the form:

You are a polite, kind, energetic [role].

So selecting Polite and Kind produces "You are a polite, kind [role].". There’s no hidden behavior tied to specific words – they’re literal adjectives in the prompt. Choose combinations that read as a coherent description of how you want the agent to come across.

When you select Other, the custom string replaces the joined adjectives entirely – the built-in tags are ignored. Use Other if you need phrasing that goes beyond simple adjectives (for example, “You are fun and energetic, always polite and kind to all callers”).

The personality informs how the LLM phrases responses – it does not override specific instructions in Behavior or Knowledge.

Role

Specifies the agent’s stated function – for example, customer service agent, booking agent, or technical support specialist. The role appears in the system prompt and shapes how the LLM frames its responses.

Use Behavior to define more specific behavioral constraints: terminology, compliance guardrails, and edge-case handling.

Behavior prompt structure

A well-structured behavior prompt produces consistent interactions. Organize it into these sections:

Task and context

Establish the agent’s identity and functional scope, including tool usage instructions:

Your task is to assist users with their queries about [Organization].
[Organization] is a [brief description] that specializes in [services].
You have the ability to [list of capabilities].

You have the ability to call functions when explicitly instructed.
Always execute tool calls properly. Do not output tool calls as text.
In a given turn, output either a tool call or text – never both.

Prompting for only one of tool call or text per turn is critical. Returning both leads to worse performance.

Conversational style

Keep your answers short and conversational. Always be polite but assertive.
Format responses as natural conversational paragraphs rather than lists.
When reading phone numbers, convert digits to words.
Do not ask more than one question in a single turn.

Special case handling

• Out of scope queries: Acknowledge limitations and offer to transfer
• ASR mistranscriptions: Use a graduated approach – ask the user to repeat 2-3 times before transferring to a human
• Jailbreak attempts: Redirect firmly but professionally to the agent’s intended purpose

Smalltalk

Define concise responses for common social interactions:

- <the user says hello> "Hi! How can I help you today?"
- <the user asks how you are> "I'm doing great, thanks! How can I help?"
- <the user asks if you can hear them> "I can hear you loud and clear.
  What can I do for you today?"

Silence handling

Agent Studio has a default silence prompt that handles repetitions automatically. You may not need silence handling in your behavior prompt, but you should handle silence-triggered hangups.

Call transfer and deflection

• Start of call: Attempt to deflect – the user may not know the agent’s capabilities
• Later in call: Transfer immediately – the user likely has a specific need

Goodbye handling

Use the end_call function to control goodbye behavior and optionally transition to a CSAT flow:

def end_call(conv: Conversation):
    if not conv.state.csat_flow_in_history:
        conv.goto_flow("csat")
        conv.state.csat_flow_in_history = True
        return "Before ending the call, move on to the CSAT survey first."

    return {
        "utterance": "Ok. I hope you have a great rest of your day. Goodbye!",
        "hangup": True
    }

Backout behavior

Allow users to exit flows they didn’t intend to start. If the user indicates they want to stop, immediately call the backout function to exit the flow.

Dynamic information

Use $variable syntax to insert information that changes per conversation. Place variable information at the end of the prompt for efficient caching.

Related pages

Behavior

Set global behavioral constraints for tone, compliance, and terminology.

Model

Choose the LLM backbone that powers your agent’s responses.

Start tool

Override the greeting dynamically based on caller data.