Skip to main content

Agents

Identity and persona

Name, use case, and the identity text the model acts from.

What identity means

A Finn has two layers of identity. The first is administrative: the name, use case, and icon you use to find the agent in your dashboard. The caller never encounters any of it. The second is the identity text, which is the opening block of the prompt the model acts from. That one the caller hears in every sentence.

Both live in the Create Finn wizard. Name, Use Case, and Finn Icon sit on the Identity tab. Identity Text sits on the Behavior tab. There is no API endpoint for editing these fields today, so all of it is done in the dashboard.

The administrative fields

FieldTabWhat it controlsCaller-visible
NameIdentityLabel in the dashboard and Finn listsNo
VoiceIdentityWhich TTS voice speaks. See voiceYes
LanguageIdentitySpeech recognition, transcript output, TTS voice constraintsYes
Language GroupIdentityVoice and STT optimization. Defaults to globalIndirectly
Time ZoneIdentityScheduled deployments and calling-hour windowsNo
Use CaseIdentityAnalytics grouping and some default behaviorsNo
Finn IconIdentityAvatar in dashboard lists. CosmeticNo
Identity TextBehaviorWho the agent is, in the model's own instructionsYes

Name

Pick something you can find later. Inbound — Demo Booking and Outbound — Renewal Reminder beat Finn 1 and test copy 2. Once you have twenty agents and each one has three deployments, a name that only describes the channel is not enough to identify anything.

Use Case

A high-level category: sales, support, scheduling, surveys. It groups your analytics and sets a small number of default behaviors. Pick the closest fit. It is not a routing decision and changing it later does not migrate historical calls into a new group.

Time Zone

Set this to where your callers live, not where you sit. It drives scheduled deployments and calling-hour windows. A time zone set to your own office is the most common cause of an outbound campaign dialing people at 3 a.m.

Language

This drives speech recognition, transcript output, and which voices are available. Pick the language your audience actually speaks rather than the language on your marketing site. If your callers switch languages mid-call, see multilingual before setting this.

Identity text

Identity text is the first thing in the prompt. It anchors four things: a name, a role, an employer, and a tone.

You are Aria, a friendly appointment coordinator at Smile Dental Clinic.
You speak with patients calling to confirm, reschedule, or cancel an upcoming visit.
You are warm, brief, and never push extra services.

Three lines. That is the target length. Identity text is not the place for company history, product catalogs, or a list of every question the agent might face. Facts belong in the knowledge base. Rules belong in Style and Guardrails, covered in guardrails.

Why length matters here specifically

Identity text sits at the front of every prompt on every turn. A multi-paragraph identity bloats the first token and slows the response, and the effect compounds across a whole call. Beyond that, a long identity dilutes itself. When you write eight paragraphs about who the agent is, no single sentence carries weight, and the model averages across all of it.

What vague identity does

"You are a helpful assistant for our company" produces an agent that will answer to any name, claim any role, and adopt whatever tone the caller sets. If a caller asks "who am I speaking to," the model improvises. If a caller pushes, the model drifts. Specific identity is what the model falls back on when the conversation goes somewhere the rules did not anticipate.

Give it a concrete tone instead of an adjective like "helpful." Three tone words work: warm, brief, never pushy. "Be professional" means nothing the model can act on.

Disclosure

If your agent may be asked whether it is human, put the answer in identity text or guardrails rather than leaving it to the model. A worked example:

If asked whether you are a real person, say honestly:
"I'm Aria, a voice assistant for Smile Dental."

Disclosure requirements vary by jurisdiction and by whether the call is recorded. See recording consent and compliance.

Editing later

Every field on this page is editable. Sidebar → View All Finn → click the Finn → Edit. Changes apply to new calls only. Calls already in progress finish under the configuration they started with, so a bad identity edit will not stop mid-call but will affect everything dialed after you save.

The platform keeps a history of configurations at View All Finn → [Finn] → History, so an identity change that degrades performance can be rolled back. See versioning.

Next

  • Prompting — the goal, rules, and examples that follow identity text.
  • Guardrails — what the agent must never say or do.
  • Voice — picking a voice that matches the identity you wrote.
  • Testing — check the identity holds under a hostile or confused caller.