Multi-Agent-Auto-Coder

(Click on Gif below to open in a new tab)

A modular, agent-based AI framework that can plan, write, run, debug, and improve code automatically — or even tackle non-coding tasks. It was originally built to work with local LLMs using Ollama, but can also use cloud-based models like OpenAI’s.

---

Features

• Task-to-code pipeline: Breaks down tasks, writes code, tests it, and fixes problems in a loop
• Project scaffolding: Builds and names project folders using LLMs
• Execution feedback: Runs code in a temporary virtual environment and tries to correct errors automatically
• Modular agents:
  • Orchestrator: Breaks down goals into steps
  • SubAgent: Handles task execution
  • Architect: Builds the project layout
  • CodeReviewer: Suggests improvements or fixes
  • Consultant: Gives high-level strategy advice
  • Overseer: Logs and approves results
  • Coder: Writes the code

CLI options like --dry-run and --no-cleanup let you preview or persist results.

---

Requirements

• Python 3.9+
• One of the following:
  • Ollama with models like llama3, codestral, etc.
  • OpenAI API access (e.g. GPT-4 or GPT-3.5)

Running with Ollama

Make sure the models are installed:

ollama run llama3
ollama run codestral

Using OpenAI

Set your API key:

export OPENAI_API_KEY=your-key-here

---

Model Configuration

Each agent is configurable through config.py, and you can use any local or API-based model that’s compatible. Set these using environment variables:

# config.py

import os

ORCHESTRATOR_MODEL = os.getenv('ORCHESTRATOR_MODEL', 'llama3.1:latest')
SUBAGENT_MODEL     = os.getenv('SUBAGENT_MODEL', 'llama3.1:latest')
REFINER_MODEL      = os.getenv('REFINER_MODEL', 'llama3.1:latest')
CONSULTANT_MODEL   = os.getenv('CONSULTANT_MODEL', 'llama3.1:latest')
BUGFIXER_MODEL     = os.getenv('BUGFIXER_MODEL', 'llama3.1:latest')
CODER_MODEL        = os.getenv('CODER_MODEL', 'llama3.1:latest')

OLLAMA_HOST = os.getenv('OLLAMA_HOST', 'http://localhost:11434')

Example:

export ORCHESTRATOR_MODEL="gpt-4"
export CODER_MODEL="codestral"

---

Usage

Interactive mode:

python main.py

With CLI flags:

python main.py --objective "Build a CSV parser" --dry-run

---

Project Layout

agentic_toolset/
├── core/
│   ├── project_manager.py
│   └── venv_manager.py
├── agents/
│   ├── reviewer.py
│   ├── architect.py
│   └── ...
├── utils/
│   └── display.py
├── config.py
main.py

---

Example Tasks (Simple)

• Create a project that calculates and logs Fibonacci numbers up to 100.
• Create a webscraper
• Make an audio file converter that converts mp3 to wav

---

Roadmap

• Command-line improvements
• Dry-run and cleanup flags
• Self-updating agent toolchain
• Integrated memory (local or vector-based)
• Optional web interface (Gradio, Textual)

---

Open to feedback, contributions, and ideas.