Agents

Overview

Agents are the core building blocks of Viber. Each agent is a specialized entity designed to perform specific tasks, with its own role, tools, and configuration.

Creating an Agent

import { Agent } from 'viber';

const writer = new Agent({
  name: 'Writer',
  model: 'openai:gpt-4o',
  systemPrompt: `You are a professional writer who creates
                 high-quality, engaging content.`,
});

Agent Configuration

Property	Type	Description
name	string	Unique identifier for the agent
model	string	LLM model identifier (e.g., openai:gpt-4o)
systemPrompt	string	Instructions defining the agent’s behavior
tools	Tool[]	Array of tools available to the agent
temperature	number	Creativity level (0-1)

Streaming Responses

const result = await agent.streamText({
  messages: [
    { role: 'user', content: 'Write a haiku about coding' }
  ],
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

Multi-Agent Collaboration

Tip

Agents can hand off tasks to each other using natural language, enabling complex workflows.

const team = [writer, reviewer];

// The lead agent coordinates the team
const lead = new Agent({
  name: 'Lead',
  model: 'openai:gpt-4o',
  systemPrompt: 'Coordinate the writer and reviewer to produce quality content.',
  team,
});

Best Practices

1. Single Responsibility: Each agent should have one clear purpose
2. Clear Prompts: Write explicit system prompts that define behavior
3. Appropriate Tools: Only provide tools relevant to the agent’s role
4. Temperature Tuning: Use lower temperatures for factual tasks, higher for creative ones