Personality
A personality defines how your agent thinks and speaks. It holds the instructions that shape every response, the messages the agent sends at the start and end of a conversation, and the choice of underlying AI model. You can create multiple personalities and assign any one of them to an agent.
What a personality contains
| Setting | Description |
|---|---|
| Instructions | What the agent knows and how it should behave — tone, rules, what to focus on, what to avoid |
| Greeting message | The first message the agent sends when a conversation starts |
| Fallback message | What the agent says if it cannot help with a request |
| Contact methods | Phone numbers, emails, URLs, and addresses the agent may share, each with a release policy |
| Underlying AI model | Which Claude model powers the agent |
Choosing an AI model
Elaras supports models from Anthropic and OpenAI.
| Model | Best for |
|---|---|
| Claude Haiku | Fast responses, lower credit cost, high-volume agents, simple Q&A |
| Claude Sonnet | More capable reasoning, complex queries, nuanced instructions |
| Claude Opus | Most capable — complex instructions, long documents, multi-step reasoning |
| GPT-4o mini | OpenAI alternative, fast and cost-efficient |
| GPT-4o | OpenAI's most capable model, comparable to Sonnet |
If you are unsure, start with Claude Haiku and switch up if you need more thorough responses.
Contact methods
Contact methods tell the agent what phone numbers, email addresses, URLs, and physical addresses it may share, and under what circumstances it may share them. Rather than writing release rules in prose instructions — where they can be ambiguous or ignored — you define each contact detail once with a precise release policy.
Each entry has four fields:
| Field | Description |
|---|---|
| Type | phone, email, url, or address |
| Usage | When the agent may give out this detail (see below) |
| Value | The actual phone number, address, link, or email |
| Primary | Mark one entry as the primary call to action (optional) |
Release policies
| Usage | When the agent shares it |
|---|---|
always | In any conversation, unprompted if relevant |
on_request | Only when the visitor explicitly asks for contact details |
fallback | Only when the agent cannot answer the visitor's question |
complaint | Only when the visitor is making a formal complaint |
custom | On a condition you describe in the Note field |
Primary call to action
Mark one entry as primary to designate it as the single most important next step — a booking link, application form, or sales enquiry URL. When the visitor shows interest in a product or in taking action, the agent offers the primary entry and nothing else alongside it. This prevents the agent from listing every contact detail at once when a focused CTA is what converts.
Example setup
A recruitment agent might have:
- A careers page URL —
usage: always, marked as primary - A phone number for the HR team —
usage: on_request - A complaints email —
usage: complaint
The agent will surface the careers link whenever a candidate shows interest, give the phone number only if asked directly, and hold the complaints email back unless a formal complaint is raised.
Versions
Personalities are versioned. You always work on a draft version. When the draft is ready, you publish it — the published version becomes the one your agent uses. Published versions are locked and cannot be edited.
| State | Behaviour |
|---|---|
| Draft | Editable. Not visible to visitors. Available in the preview panel for testing. |
| Published | Locked. Active on any agents assigned to this personality. |
Only one version can be published at a time. Publishing a new version automatically replaces the previous one.
Creating a personality
Go to Personality > New personality in your agent's settings. Give it a name, then configure the first version:
- Write the instructions
- Set the greeting and fallback messages
- Choose the underlying model
- Save as draft
Editing a draft
Draft versions are fully editable. Open the draft and make your changes — instructions, greeting, model choice. Changes save automatically.
Publishing a version
Before you can publish, the draft must pass the testing gate: at least one scenario must exist and the most recent test run must pass. See Testing for details.
When ready, click Publish. Agents assigned to this personality switch to the new version immediately.
Cloning a version
To iterate on a published version without editing it directly, open the version menu and choose Clone. This creates a new draft that is an exact copy. Use it for instruction refinements, A/B testing, or rolling back — clone the previous published version, then publish the clone.
Assigning a personality to an agent
An agent uses one published personality at a time. Open the agent's Settings tab and choose a personality from the dropdown. The agent immediately starts using that personality's currently published version. When you publish a new version later, the agent picks it up automatically — no reassignment needed.
Typical workflow
- Clone the current published version.
- Edit the instructions in the new draft.
- Test using the preview panel and run scenarios.
- Publish the draft — all assigned agents switch over immediately.
- If something is wrong, clone the previous published version and publish it to roll back.
