Agentflow

Agentflow is a low-code framework for creating and executing AI-powered workflows using Markdown and natural language.

• 💬 Natural language as code - Write powerful workflows using natural language and Markdown, making automation accessible to all.
• 🧠 AI-powered execution - Take advantage of cutting-edge AI models to bring super-intelligence to your workflows.
• 🔌 Flexible and adaptable - Connect with any AI provider or run models locally, for complete control and flexibility.
• 🔀 Full logical control - Create complex workflows with loops and conditional branching, just like traditional programming but in plain English.
• 🛠️ Extend with ease - Add custom actions and tools using JavaScript to tailor Agentflow to your specific needs.

Warning

Agentflow is in early development. While it's ready for experimentation and early adoption, the API may change as we continue to improve the framework. Always refer to the documentation for the latest updates.

Table of contents

• Quick start
• Writing workflows
• JavaScript API
• Documentation
• License

Quick start

Agentflow requires a JavaScript runtime (for example, Node.js) to be installed on your system. Once you have that set up, there are two simple ways to get started with Agentflow: using the starter kit to bootstrap a project, or using the CLI to initialize a project manually.

1. Using the Agentflow starter kit

The quickest way to get started is using the Agentflow starter kit. Choose your preferred package manager:

# Using NPM
npm create agentflow@latest my-agents

# Using Yarn
yarn create agentflow my-agents

This will scaffold a new project with all the necessary dependencies and configuration. The CLI will be installed locally as a project dependency, which you can use via your package manager:

# Using NPM
cd my-agents && npm install
npx aflow help

# Using Yarn
cd my-agents && yarn
yarn aflow help

2. Using the global CLI

Alternatively, you can install the Agentflow CLI globally and initialize a new project via the CLI:

# Install CLI globally
npm install -g @agentflow/cli

# Create new project
aflow init my-agents
cd my-agents

For more detailed instructions and configuration options, check out the getting started guide and CLI documentation.

Writing workflows

Workflows in Agentflow are written using Markdown files that combine natural language instructions with actions using MDX-like syntax. Actions can do anything from generating text with AI models to implementing control flow and integrating external services.

Here's a simple example:

---
data:
  languages:
    - Spanish
    - French
    - German
input:
  topic:
    type: text
    message: "Enter a topic to write about"
---

Write a short blog post about {topic}.

<GenText as="article" model="openai:gpt-4o" />

<Loop
  as="translations"
  until={$.index === languages.length}
  provide={{
    original: article,
    language: languages[$.index],
  }}>

  Translate this article to {language}:
  {original}

  <GenText as="translated" model="openai:gpt-4o" />
</Loop>

This workflow prompts for a topic, generates a blog post, then automatically translates it into multiple languages. Actions like <GenText /> invoke AI models to generate content, while control flow actions like <Loop /> enable complex workflows. Expressions in braces (e.g. {topic}) reference previous inputs and results.

Check out the docs to learn more about workflow structure and working with input data.

JavaScript API

Agentflow provides a JavaScript API for working with workflows programmatically. Using @agentflow/core, you can compile Markdown-formatted workflows and control their execution from your code:

import { Environment, Workflow } from '@agentflow/core'

// Initialize the environment
const env = new Environment({
  providers: {
    // configure AI providers
  }
})

// Compile a workflow from markdown
const workflow = Workflow.compileSync(`
  Write a haiku about {topic}.

  <GenText as="haiku" model="openai:gpt-4o-mini" />
`, env)

// Create and run the execution
const ctrl = workflow.createExecution({
  topic: { type: 'primitive', value: 'spring rain' }
})

// Handle execution events
ctrl.on('step', (step, event) => {
  event.action?.then(({ result }) => console.log(result))
})

ctrl.on('complete', (output) => {
  console.log('Done!\n', output)
})

// Run the workflow
await ctrl.runAll()

The API provides fine-grained control over workflow execution, including the ability to pause, resume and monitor progress. Check out the JavaScript API documentation for more details.

Documentation

For more detailed guides, examples and API documentation, visit agentflow.2point0.ai.

License

This package is open source and released under the Apache-2 Licence.

© Copyright 2024 Push Code Ltd.