Voice and Persona

Your agent's voice and persona determine how it sounds and behaves. A well-configured persona produces more accurate, on-brand responses. A well-chosen voice makes the agent feel like a natural participant in the conversation.

→ Open Agent Builder →

Persona Prompt#

The persona prompt is the AI's system context — a brief description of who the agent is, what it knows, and how it should behave. It's included in every evaluation and shapes every response the agent produces.

Why it matters#

Without a persona prompt, the agent responds generically. With a good persona prompt, it knows:

• What company or team it belongs to
• What its area of expertise is
• What tone to take (formal, friendly, concise, detailed)
• What it should and should not say
• Any specific knowledge it should assume or reference

Writing a strong persona prompt#

Keep it to 2–4 short paragraphs. Longer prompts don't always produce better results — and they consume more of the evaluation context window.

Be specific about role and knowledge:

Set tone expectations:

Add guardrails if needed:

Voice Selection (Live agents only)#

Live agents speak aloud during meetings using AI voices powered by ElevenLabs. CoAgentor provides a curated library of voices to choose from.

Choosing a voice#

In the Agent Builder, click Choose Voice to open the voice selector. Each voice has:

• A display name
• A brief description (e.g. "Professional male voice, American English")
• A Play button to hear a preview

Select the voice that best fits the agent's persona — a formal executive coach sounds different from a friendly sales assistant.

Voice characteristics to consider#

• Accent and dialect — currently English voices with American and British accents are available
• Pacing — some voices naturally speak faster or slower; consider your meeting context
• Formality — some voices sound more authoritative, others more conversational
• Gender presentation — choose what fits your brand or scenario

Multi-language voices#

Support for voices in languages other than English is coming in a future release. If your team conducts meetings in multiple languages, stay tuned — agent voice, transcription, and response generation will all support locale configuration.

→ Learn about multi-language support (coming soon) →

How speech is processed#

When a trigger fires on a Live agent, CoAgentor:

1. Generates the response text using the AI evaluation engine
2. Passes the text through a speech normalisation layer — this converts shorthand like $31K, &, and Q3 into forms that sound natural when spoken (e.g. "thirty-one thousand dollars", "and", "Q3")
3. Sends the normalised text to ElevenLabs for synthesis
4. Streams the resulting audio back into the meeting in real time via the Recall.ai audio injection pipeline

The original un-normalised text is always what's stored in your meeting event log and shown in summaries — you always see the agent's exact intended response, not the TTS-adjusted version.

Tips for persona prompts#

✅ Specify the company or product context — the agent should know where it works

✅ Set a maximum response length instruction — "no more than 3 sentences" dramatically improves meeting flow

✅ Tell it what not to do — negative instructions are often the most valuable guardrails

✅ Include sample phrasing if you have brand voice guidelines — "use phrases like X, avoid phrases like Y"

❌ Don't write a novel — 4 paragraphs is usually enough

❌ Don't ask the agent to do things that trigger rules already control — the persona sets character, trigger rules set behaviour

❌ Don't include information the agent will never need — keep the context focused on the meeting scenarios it will actually encounter