Workflow Agents Getting Started - Upstash Documentation

In this guide, we will be using Next.js. If you’re working with a different supported framework, you can find instructions on how to define a workflow endpoint in the quickstarts. If you’re new to Upstash Workflow, it’s a good idea to start by exploring the Local Development documentation. This guide will help you set up and use Upstash Workflow in a local environment.

Installation

First, create a new Next.js project with:

npx create-next-app@latest [project-name] [options]

Then, install the following packages:

npm i @upstash/workflow @upstash/workflow-agents ai zod

Start the development server

Since you are using @upstash/workflow, set QSTASH_DEV=true in your .env.local file and the SDK will download, start, and connect to a local QStash development server automatically — no tokens or signing keys to copy over. Add your OPENAI_API_KEY alongside it:

.env.local

QSTASH_DEV=true

OPENAI_API_KEY=<OPENAI_API_KEY>

For manual setup, custom ports, and the registerQStashDev() helper for Next.js edge routes, see the Development Server guide.

Define an endpoint

Next, we will define the endpoint to run the agent.

app/workflow/route.ts

import { z } from "zod";
import { tool } from "ai";

import { serve } from "@upstash/workflow/nextjs";
import { agentWorkflow } from "@upstash/workflow-agents";

export const { POST } = serve<{ prompt: string }>(async (context) => {
  const prompt = context.requestPayload.prompt
  const agents = agentWorkflow(context)
  const model = agents.openai('gpt-3.5-turbo')

  const communicatorAgent = agents.agent({
    model,
    name: 'communicatorAgent',
    maxSteps: 2,
    tools: {
      communicationTool: tool({
        description: 'A tool for informing the caller about your inner thoughts',
        parameters: z.object({ message: z.string() }),
        execute: async ({ message }) => {
          console.log("Inner thought:", message)
          return "success"
        }
      })
    },
    background:
      'Answer questions directed towards you.' +
      ' You have access to a tool to share your inner thoughts' +
      ' with the caller. Utilize this tool at least once before' +
      ' answering the prompt. In your inner thougts, briefly' +
      ' explain what you will talk about and why. Keep your' +
      ' answers brief.',
  })

  const task = agents.task({
    agent: communicatorAgent,
    prompt
  })

  const { text } = await task.run()

  console.log("Final response:", text);
})

You can refer to the documentations for defining a workflow endpoint and the Agents API features to learn more.

Calling the Endpoint

To run the endpoint, first run the Next.js app with:

npm run dev

Then, we call the endpoint using the Workflow Client:

import { Client } from "@upstash/workflow";

const client = new Client({
  baseUrl: process.env.QSTASH_URL,
  token: process.env.QSTASH_TOKEN!,
})

const workflowRunId = await client.trigger({
  url: "http://127.0.0.1:3000/workflow",
  body: { prompt: "Explain the future of space exploration" }
})

console.log(workflowRunId);

If you are using a local tunnel, replace the url above (http://127.0.0.1:3000) with the public URL.

In the console where you run the Next.js app, you should see logs like this:

Inner thought: I will discuss the future of space
exploration and the potential advancements in
technology and missions.

Final response: The future of space exploration
holds exciting possibilities with advancements
in technology, potential manned missions to
Mars, increased commercial space travel,
and exploration of distant celestial
bodies.

If you run the same endpoint using a local tunnel, you can also see how Upstash Workflow runs the agent in steps:

Each tool invocation and LLM call is a seperate step. Our agent first made a call to OpenAI to decide whether to use a tool or reply right away. OpenAI responded with a request to use the tool communicationTool. Tool was executed and OpenAI was called with the result of the tool. OpenAI then responded with the final response.

​Installation

​Start the development server

​Define an endpoint

​Calling the Endpoint

Installation

Start the development server

Define an endpoint

Calling the Endpoint