Create an agent from scratch

Let's create an agent from scratch. We are inside a project, and here, under Agents, we can see the agents it already has.

To create a new one, click Create. Then the Create New Agent page opens.

At the top, you can start from scratch, with full control over your agent's persona, behavior and rules. And further down, you will find a library of ready made templates, which you can filter by industry and by use case. A template fills in the form for you.

We want full control, so we click Start from Scratch. The form opens, starting with the agent's profile.

First, the agent type. A relational agent uses advanced voice capabilities and parallel thinking, which is perfect for complex use cases and deep interactions. A lightweight agent uses efficient voices with a smaller reasoning footprint, which keeps the cost much lower, while staying reliable.

For this agent, we will use a relational agent. Keep in mind that the type cannot be changed after the agent is saved.

Next, give your agent a name, a job title, and a short description. People see these in the app, but they do not change how the agent behaves. Let's call this one Mia, a welcome guide.

On the right, you can see a preview of your agent, which reflects the choices you make.

Now the agentic purpose. The objective tells the agent what it should achieve. Keep it practical, and describe what success looks like. Here, we ask Mia to welcome new visitors, and to help them find the right next step.

Guardrails are the rules your agent must follow. What it should do, and what it must never do. For example, Mia must never give legal, medical or financial advice.

Business process steps describe the ideal path of the conversation, step by step, like a script. This one is only for relational agents. Here, Mia greets the visitor, explains the answer, offers a next step, and closes politely.

The relational settings slider decides how closely the agent follows those steps. Move it toward relational focus for a more conversational agent, or toward task focus, to follow the steps tightly. Balanced is the default.

And remembers users decides whether the agent remembers the same person between conversations. It is on by default. Turn it off, and every conversation starts fresh.

Now the agent's identity, which is how your agent looks. The avatar source decides how the avatar is shown, either rendered on the device, or streamed from a LiveAvatar account.

Then pick the avatar itself. You can filter the list by gender, and the first option is an animated orb.

Then choose an environment, which is the visual background behind your agent, and a camera preset. The camera preset is static by default, which keeps a steady framing of your avatar.

Next, choose the model that powers your agent, and its voice. You can filter the voices by language and by gender.

Then pick the allowed languages. These are the languages your agent can speak. You can pick one or more, and they depend on the voice you chose.

The character and backstory describes who your agent is, their personality, their tone of voice, and how they behave in tricky moments. We make Mia warm, patient and clear.

Finally, the user's evaluation. Here you can attach evaluations that assess the user's side of the conversation, not the agent. You can search for an evaluation, or create a new one, and they run during the conversation, or after it ends.

And that is everything. Click Save, and your agent is created. It only takes a moment.

Your agent is created, and ready to have a great conversation.