AI Agent Flow

Why We Built It

Every AI product we build — AIVisibility, ProfitLab, MyJob.fit — needed the same skeleton: call a model, branch on the result, call a tool, maybe loop, maybe hand off to a human. Each time we wrote it by hand, slightly differently, slightly worse.

So we extracted the pattern into a library, then built a visual editor on top, then open-sourced the whole thing. If we need it in every product, someone else does too.

What It Does

AI Agent Flow lets you define agent workflows as composable nodes:

• Model nodes — LLM calls with any provider
• Tool nodes — functions the agent can invoke
• Branch nodes — conditional routing on structured output
• Loop nodes — retry and refinement cycles
• Human nodes — pause for approval or input

Workflows are defined in code (type-safe) or visually (the builder exports the same definition). Run them locally, in a serverless function, or anywhere JavaScript runs.

Installation

npm install aiagentflow

Usage

import { workflow, model, branch } from 'aiagentflow';

const support = workflow()
  .then(model('classify-intent'))
  .then(branch({
    billing: model('resolve-billing'),
    bug: model('file-bug-report'),
  }));

const result = await support.run({ message: 'My card was charged twice' });

Contributing

Issues and pull requests welcome at the repository. The project follows conventional commits and requires types and tests for new nodes.

Current State

Actively maintained, MIT-licensed, and used in our own AI products — the best motivation to keep it good.