AgentCreator

An AutoGen multi-agent system that generates business ideas and scaffolds domain-expert agents to refine them.

On each run, IdeaSmith proposes a new business idea, Creator writes a specialized agent from a template, registers it on the runtime, and that domain expert refines the idea. The result is saved as markdown in the project root.

Demo

Each run produces a new domain-expert agent under src/agent_forge/ and a refined business idea saved as idea_<domain>.md in the project root.

Generated domain agents

Creator writes and registers a specialist agent from template_agent.py. Examples from two runs:

Education technology	Fashion tech & e-commerce

Refined business ideas

The generated agent enhances the initial idea with domain-specific detail:

Education technology	Fashion tech & e-commerce

How it works

flowchart LR
    start["agent-forge"] --> ideaSmith["IdeaSmith"]
    ideaSmith --> idea["Business Idea"]
    idea --> creator["Creator"]
    creator --> agent["Generated domain agent"]
    agent --> refined["Refined idea"]
    refined --> output["idea_<domain>.md"]
    creator --> code["src/agent_forge/<agent_name>.py"]

Loading

Agent	Role
IdeaSmith	Generates a lightweight business idea (investment, returns, domain)
Creator	Writes a new agent module from template_agent.py, loads it, and runs it
Generated agent	Refines the idea using domain expertise

Prerequisites

• Python >= 3.10, < 3.14
• uv
• OpenAI API key (see Setup)

Setup

From the project root:

uv sync
cp .env.example .env
# Edit .env and set OPENAI_API_KEY

Environment variables

Variable	Required	Description
OPENAI_API_KEY	Yes	OpenAI API key for agent reasoning
OPENAI_MODEL	No	Model name (default: gpt-4o-mini)

Run

From the project root, with dependencies installed via uv sync:

uv run agent-forge

Other equivalent ways to start the pipeline:

# Using the virtualenv directly
.venv/bin/agent-forge

# As a Python module
uv run python -m agent_forge.main

The command makes several OpenAI API calls (idea generation, agent codegen, idea refinement). Progress is logged to the terminal.

Output

After a successful run:

• idea_<domain>.md — refined business idea written to the current working directory (for example, idea_education_technology_online_tutoring_services.md)
• src/agent_forge/<agent_name>.py — newly generated domain-expert agent module (for example, personalized_virtual_tutoring.py)

Run the command from the project root so output paths land where you expect.

Project layout

assets/                  # README demo screenshots
src/agent_forge/
├── main.py              # CLI entry point (`agent-forge`)
├── idea_smith.py        # Generates business ideas
├── creator.py           # Creates and registers domain agents
├── template_agent.py    # Template passed to the LLM for codegen
├── messages.py          # Shared message types (Idea, AgentSpecs, …)
└── *.py                 # Generated agent modules (created at runtime)

Tech stack

• AutoGen Core — agent runtime and messaging
• AutoGen AgentChat — assistant agents and structured outputs
• uv — dependency and project management

License

TBD