TypeScript Agent Workflow

A TypeScript framework for building and orchestrating AI agent workflows. This project provides a structured way to create multi-step workflows with AI agents, allowing for both automated processes and interactive user experiences.

Features

• Workflow Orchestration: Define and execute multi-step workflows with clear input/output schemas
• Agent Steps: Create AI-powered steps that can interact with users or run autonomously
• Type Safety: Fully typed with TypeScript for better developer experience and code reliability
• Schema Validation: Uses Zod for runtime type validation of inputs and outputs
• Flexible Input Sources: Configure steps to receive input from previous steps, global state, or user interaction

Installation

git clone https://github.com/TiagoGouvea/TypeScript-Agent-Workflow.git
cd TypeScript-Agent-Workflow
nvm install
nvm use
yarn
echo "OPENAI_API_KEY=your_api_key_goes_here" > .env

Usage

Running the Project

You can see all agencies and workflow and select one running:

yarn start

This will display a list of available agencies and workflows that you can choose to run.

Or you can run a specific workflow:

yarn start --workflow=agentTest

Creating a Workflow

Here's an example of how to create a simple workflow:

import { z } from 'zod';
import { createAgentStep, InputSource } from "../../types/workflow/Step.ts";
import { Workflow } from "../../types/workflow/Workflow.ts";

// Step 1: Create a character
const agentTest = createAgentStep({
  name: 'Create a character',
  introductionText: 'I will create a character',
  inputSchema: z.object({
    subject: z
      .string()
      .min(1, 'Title is required')
      .describe('Some details about how the character should be'),
  }),
  inputSource: InputSource.UserInput,
  outputSchema: z.object({
    name: z.string(),
    age: z.number(),
    gender: z.string(),
    description: z.string(),
    traits: z.array(z.string()),
    backstory: z.string(),
    goal: z.string(),
    objectives: z.array(z.string()),
  }),
  systemPrompt: `Create a character`,
});

// Step 2: Write a story with the character
const writeStory = createAgentStep({
  name: 'Write a story',
  introductionText: 'I will write a story',
  inputSchema: agentTest.outputSchema,
  inputSource: InputSource.LastStep,
  outputSchema: z
    .string()
    .describe('A one-paragraph story with the character'),
  systemPrompt: `Create a story`,
});

// Create the workflow with the defined steps
export const myWorkflow = new Workflow({ agentTest, writeStory });

Project Structure

• src/types/workflow/: Core workflow types and classes
• src/_workflows/: Workflow definitions
• src/_agencies/: Agency definitions
• src/utils/: Utility functions
• src/llm/: LLM integration code

Exemples

• /src/_workflows/createCharacter/ Simpler - Two steps workflow with user input on the firs step, with LLM interaction on both steps.
• /src/_workflows/dailyNews/ Simpler - Two steps workflow with user input on the firs step, with LLM interaction on both steps.

Architecture

The framework is built around these key concepts:

1. Step: A single unit of work with defined input and output schemas
2. Workflow: A collection of steps that are executed in sequence
3. InputSource: Defines where a step gets its input (previous step, user input, etc.)
4. Agent Step: A specialized step that uses an AI model to process inputs and generate outputs

License

ISC

Author

Tiago Gouvêa