Logo
Modules/Agents

Agent Workflows

Agent Workflows are a powerful system that enables you to create and orchestrate one or multiple agents with tools to perform specific tasks. It's built on top of the base Workflow system and provides a streamlined interface for agent interactions.

Usage

Single Agent Workflow

The simplest use case is creating a single agent with specific tools. Here's an example of creating an assistant that tells jokes:

import { tool } from "llamaindex";
import { agent } from "@llamaindex/workflow";
import { openai } from "@llamaindex/openai";

// Define a joke-telling tool
const jokeTool = tool(
  () => "Baby Llama is called cria",
  {
    name: "joke",
    description: "Use this tool to get a joke",
  }
);

// Create an single agent workflow with the tool
const jokeAgent = agent({
  tools: [jokeTool],
  llm: openai({ model: "gpt-4o-mini" }),
});

// Run the workflow
const result = await jokeAgent.run("Tell me something funny");
console.log(result.data.result); // Baby Llama is called cria
console.log(result.data.message); // { role: 'assistant', content: 'Baby Llama is called cria' }

Structured Output

You can extract structured data from agent responses by providing a responseFormat with a Zod schema. This is useful when you need the agent's response in a specific format for further processing:

import { z } from "zod";
import { tool } from "llamaindex";
import { agent } from "@llamaindex/workflow";
import { openai } from "@llamaindex/openai";

// Define a weather tool
const weatherTool = tool({
  name: "weatherTool",
  description: "Get weather information",
  parameters: z.object({
    location: z.string(),
  }),
  execute: ({ location }) => {
    return `The weather in ${location} is sunny. The temperature is 72 degrees. The humidity is 50%. The wind speed is 10 mph.`;
  },
});

// Define the structure you want for the response
const responseSchema = z.object({
  temperature: z.number(),
  humidity: z.number(),
  windSpeed: z.number(),
});

// Create the agent
const weatherAgent = agent({
  name: "weatherAgent",
  tools: [weatherTool],
  llm: openai({ model: "gpt-4.1-mini" }),
});

// Run with structured output
const result = await weatherAgent.run("What's the weather in Tokyo?", {
  responseFormat: responseSchema,
});

console.log("Natural language result:", result.data.result);
console.log("Structured data:", result.data.object);
// Output: { temperature: 72, humidity: 50, windSpeed: 10 }

The agent will:

  1. Use the weather tool to get the raw weather information
  2. Process that information through the LLM
  3. Extract structured data according to your schema
  4. Return both the natural language response and the structured object

Event Streaming

Agent Workflows provide a unified interface for event streaming, making it easy to track and respond to different events during execution:

import { agentToolCallEvent, agentStreamEvent } from "@llamaindex/workflow";

// Get the workflow execution context
const events = jokeAgent.runStream("Tell me something funny");

// Stream and handle events
for await (const event of events) {
  if (agentToolCallEvent.include(event)) {
    console.log(`Tool being called: ${event.data.toolName}`);
  }
  if (agentStreamEvent.include(event)) {
    process.stdout.write(event.data.delta);
  }
}

Multi-Agent Workflow

An Agent Workflow can orchestrate multiple agents, enabling complex interactions and task handoffs. Each agent in a multi-agent workflow requires:

  • name: Unique identifier for the agent
  • description: Purpose description used for task routing
  • tools: Array of tools the agent can use
  • canHandoffTo (optional): Array of agent names or agent instances that this agent can delegate tasks to

Here's an example of a multi-agent system that combines joke-telling and weather information:

import { tool } from "llamaindex";
import { multiAgent, agent } from "@llamaindex/workflow";
import { openai } from "@llamaindex/openai";
import { z } from "zod";

// Create a weather agent
const weatherAgent = agent({
  name: "WeatherAgent",
  description: "Provides weather information for any city",
  tools: [
    tool(
      {
        name: "fetchWeather",
        description: "Get weather information for a city",
        parameters: z.object({
          city: z.string(),
        }),
        execute: ({ city }) => `The weather in ${city} is sunny`,
      }
    ),
  ],
  llm: openai({ model: "gpt-4o-mini" }),
});

// Create a joke-telling agent
const jokeAgent = agent({
  name: "JokeAgent",
  description: "Tells jokes and funny stories",
  tools: [jokeTool], // Using the joke tool defined earlier
  llm: openai({ model: "gpt-4o-mini" }),
  canHandoffTo: [weatherAgent], // Can hand off to the weather agent
});

// Create the multi-agent workflow
const agents = multiAgent({
  agents: [jokeAgent, weatherAgent],
  rootAgent: jokeAgent, // Start with the joke agent
});

// Run the workflow
const result = await agents.run(
  "Give me a morning greeting with a joke and the weather in San Francisco"
);
console.log(result.data.result);

The workflow will coordinate between agents, allowing them to handle different aspects of the request and hand off tasks when appropriate.

Edit on GitHub

Last updated on