Workflow Orchestrator Demo

This is a demonstration of using AIGNE Framework to build an orchestrator workflow. The example now supports both one-shot and interactive chat modes, along with customizable model settings and pipeline input/output.

flowchart LR

in(In)
out(Out)
orchestrator(Orchestrator)
synthesizer(Synthesizer)
finder(Finder)
writer(Writer)
proofreader(Proofreader)
fact_checker(Fact Checker)
style_enforcer(Style Enforcer)

in ==> orchestrator
orchestrator -.-> finder -.-> synthesizer
orchestrator -.-> writer -.-> synthesizer
orchestrator -.-> proofreader -.-> synthesizer
orchestrator -.-> fact_checker -.-> synthesizer
orchestrator -.-> style_enforcer -.-> synthesizer
synthesizer ==> out

classDef inputOutput fill:#f9f0ed,stroke:#debbae,stroke-width:2px,color:#b35b39,font-weight:bolder;
classDef processing fill:#F0F4EB,stroke:#C2D7A7,stroke-width:2px,color:#6B8F3C,font-weight:bolder;

class in inputOutput
class out inputOutput
class orchestrator processing
class synthesizer processing
class finder processing
class writer processing
class proofreader processing
class fact_checker processing
class style_enforcer processing

Prerequisites

• Node.js (>=20.0) and npm installed on your machine
• An OpenAI API key for interacting with OpenAI's services
• Optional dependencies (if running the example from source code):
  • Bun for running unit tests & examples
  • Pnpm for package management

Quick Start (No Installation Required)

export OPENAI_API_KEY=YOUR_OPENAI_API_KEY # Set your OpenAI API key

# Run in one-shot mode (default)
npx -y @aigne/example-workflow-orchestrator

# Run in interactive chat mode
npx -y @aigne/example-workflow-orchestrator --chat

# Use pipeline input
echo "Research ArcBlock and compile a report about their products and architecture" | npx -y @aigne/example-workflow-orchestrator

Installation

Clone the Repository

git clone https://github.com/AIGNE-io/aigne-framework

Install Dependencies

cd aigne-framework/examples/workflow-orchestrator

pnpm install

Setup Environment Variables

Setup your OpenAI API key in the .env.local file:

OPENAI_API_KEY="" # Set your OpenAI API key here

When running Puppeteer inside a Docker container, set the following environment variable:

DOCKER_CONTAINER="true"

This ensures Puppeteer configures itself correctly for a Docker environment, preventing potential compatibility issues.

Using Different Models

You can use different AI models by setting the MODEL environment variable along with the corresponding API key. The framework supports multiple providers:

• OpenAI: MODEL="openai:gpt-4.1" with OPENAI_API_KEY
• Anthropic: MODEL="anthropic:claude-3-7-sonnet-latest" with ANTHROPIC_API_KEY
• Google Gemini: MODEL="gemini:gemini-2.0-flash" with GEMINI_API_KEY
• AWS Bedrock: MODEL="bedrock:us.amazon.nova-premier-v1:0" with AWS credentials
• DeepSeek: MODEL="deepseek:deepseek-chat" with DEEPSEEK_API_KEY
• OpenRouter: MODEL="openrouter:openai/gpt-4o" with OPEN_ROUTER_API_KEY
• xAI: MODEL="xai:grok-2-latest" with XAI_API_KEY
• Ollama: MODEL="ollama:llama3.2" with OLLAMA_DEFAULT_BASE_URL

For detailed configuration examples, please refer to the .env.local.example file in this directory.

Run the Example

pnpm start # Run in one-shot mode (default)

# Run in interactive chat mode
pnpm start -- --chat

# Use pipeline input
echo "Research ArcBlock and compile a report about their products and architecture" | pnpm start

Run Options

The example supports the following command-line parameters:

| Parameter | Description | Default | |-----------|-------------|---------| | --chat | Run in interactive chat mode | Disabled (one-shot mode) | | --model <provider[:model]> | AI model to use in format 'provider[:model]' where model is optional. Examples: 'openai' or 'openai:gpt-4o-mini' | openai | | --temperature <value> | Temperature for model generation | Provider default | | --top-p <value> | Top-p sampling value | Provider default | | --presence-penalty <value> | Presence penalty value | Provider default | | --frequency-penalty <value> | Frequency penalty value | Provider default | | --log-level <level> | Set logging level (ERROR, WARN, INFO, DEBUG, TRACE) | INFO | | --input, -i <input> | Specify input directly | None |

Examples

# Run in chat mode (interactive)
pnpm start -- --chat

# Set logging level
pnpm start -- --log-level DEBUG

# Use pipeline input
echo "Research ArcBlock and compile a report about their products and architecture" | pnpm start

Example

The following example demonstrates how to build a orchestrator workflow:

Here is the generated report for this example: arcblock-deep-research.md

import { OrchestratorAgent } from "@aigne/agent-library/orchestrator/index.js";
import { AIAgent, AIGNE, MCPAgent } from "@aigne/core";
import { OpenAIChatModel } from "@aigne/core/models/openai-chat-model.js";

const { OPENAI_API_KEY } = process.env;

const model = new OpenAIChatModel({
  apiKey: OPENAI_API_KEY,
  modelOptions: {
    parallelToolCalls: false, // puppeteer can only run one task at a time
  },
});

const puppeteer = await MCPAgent.from({
  command: "npx",
  args: ["-y", "@modelcontextprotocol/server-puppeteer"],
  env: process.env as Record<string, string>,
});

const filesystem = await MCPAgent.from({
  command: "npx",
  args: ["-y", "@modelcontextprotocol/server-filesystem", import.meta.dir],
});

const finder = AIAgent.from({
  name: "finder",
  description: "Find the closest match to a user's request",
  instructions: `You are an agent that can find information on the web.
You are tasked with finding the closest match to the user's request.
You can use puppeteer to scrape the web for information.
You can also use the filesystem to save the information you find.

Rules:
- do not use screenshot of puppeteer
- use document.body.innerText to get the text content of a page
- if you want a url to some page, you should get all link and it's title of current(home) page,
then you can use the title to search the url of the page you want to visit.
  `,
  skills: [puppeteer, filesystem],
});

const writer = AIAgent.from({
  name: "writer",
  description: "Write to the filesystem",
  instructions: `You are an agent that can write to the filesystem.
  You are tasked with taking the user's input, addressing it, and
  writing the result to disk in the appropriate location.`,
  skills: [filesystem],
});

const agent = OrchestratorAgent.from({
  skills: [finder, writer],
  maxIterations: 3,
  tasksConcurrency: 1, // puppeteer can only run one task at a time
});

const aigne = new AIGNE({ model });

const result = await aigne.invoke(
  agent,
  `\
Conduct an in-depth research on ArcBlock using only the official website\
(avoid search engines or third-party sources) and compile a detailed report saved as arcblock.md. \
The report should include comprehensive insights into the company's products \
(with detailed research findings and links), technical architecture, and future plans.`,
);
console.log(result);
// Output:
// {
//   $message: "The objective of conducting in-depth research on ArcBlock using only the official website has been successfully completed...",
// }

License

This project is licensed under the MIT License.