Skip to main content
Agents are the core of the VIVI platform. An agent is your virtual team member that interacts with customers, employees, or systems. Agents combine several components, such as knowledge bases and prompts, into one system. These components allow your agent to deliver helpful, accurate, and efficient support. At a minimum, every agent requires a name, prompt, and model to be live. Once configured, you can use the Test Agent feature or Evaluations to validate responses before deployment. Unlike traditional decision trees or scripted bots, VIVI agents are dynamic — they reason, act, and adapt to context. Agents can be duplicated, customized, or scaled as your organization evolves.

​
Building Agents

When building your agent, you can customize every detail to match your organization’s standards, tone, and goals.
1

Choose a Name & Description

Choose a clear name and description that reflects the agent’s purpose.
2

Add a Prompt

Define the agent’s role, tone, and rules of engagement.
3

Select a Model

Choose the large language model that powers the agent. See Models for a full breakdown of available options.
4

Connect Knowledge Bases

Provide the internal sources of truth the agent can reference.
5

Enable Integrations

Connect MCP servers, APIs, or even other agents for advanced actions.
6

Add Skills

Add custom built skills to your agent to allow it to run code locally.
7

Enable Citations

Allow the agent to include numbered references to knowledge base sources in its responses. See Citations for setup details.
8

Enable Structured Output

Force the agent to reply with a typed JSON object instead of free-form text. Useful for API-driven workflows. See Structured Output for details.
9

Select Channel

Decide where the agent will be active (web chat, WhatsApp, API, etc.).
10

Test the Agent

Use the testing features to verify accuracy and performance before deployment.

​
Best Practices

  • Use descriptive names so team members can quickly identify each agent’s role.
  • Choose large-context models for complex reasoning and smaller, faster models for lightweight workflows.
  • Enable citations when your agent draws from knowledge bases — it builds user trust and makes responses easier to verify.
  • Use structured output only for API-driven workflows where a downstream system needs to parse the agent’s response as JSON.
  • Test thoroughly after configuring each component to ensure accuracy and reliability.