Introducing Agentflow: Execute complex LLM workflows with simple JSON

Large language models (LLMs) are powerful tools, but implementing complex workflows with them can be a challenge.

Yes, tools like Auto-GPT and BabyAGI allow LLMs to execute multiple steps, but autonomously—the LLMs plan and then execute tasks themselves. Because of this, in my experience with Auto-GPT, things can quickly get out of control.

What I want is to have LLMs execute multiple steps, but under my control, following a predefined path. So I scratched my own itch and built Agentflow, an open source solution that lets you execute complex workflows with simple JSON.

With Agentflow, you can:

1. Write workflows in plain English

Just add tasks in a JSON file like this:

{
    "system_message": "Optional guiding message",
    "tasks": [
        {
            "action": "Step one."
        },
        {
            "action": "Step two."
        },
        {
            "action": "..."
        }
    ]
}

2. Add variables for dynamic outputs

You can include variables in {curly quotes} that you populate when running a workflow. For example, target_market is a variable here:

{
    "system_message": "You are an innovative entrepreneur.",
    "tasks": [
        {
            "action": "Generate 10 product ideas for {target_market}"
        },
        {
            "action": "..."
        }
    ]
}

3. Create and use custom functions

Custom functions expand LLMs’ capabilities beyond text generation. Easily define new functions by inheriting from the BaseFunction class. Specify functions to run using function_call as shown here:

{
    "system_message": "You are a creative artist.",
    "tasks": [
        {
            "action": "Brainstorm 10 painting ideas for {painting_subject}."
        },
        {
            "action": "Choose the best idea."
        },
        {
            "action": "Write a prompt for an AI art generator to produce an image of the painting."
        },
        {
            "action": "Generate the painting image using the prompt.", 
            "settings": {
                "function_call": "create_image"
            }
        },
        {
            "action": "..."
        }
    ]
}

4. Run workflows with a simple command

To run a workflow, just use the command line like this:

python -m run --flow=workflow_name

Or, for workflows with variables, like this:

python -m run --flow=workflow_with_variables_name --variables 'variable_1_name=value1' 'variable_2_name=value2'

Agentflow executes the specified workflow and provides a link to a folder with all outputs, including a JSON file containing all of the LLM’s responses.

Get started with Agentflow!

Check out the installation instructions, explore ideas and open issues, and feel free to contribute to expanding Agentflow’s capabilities.