📃 arXiv Paper · 📓 NeurIPS 2024 · ❞ Citation

• 🪶 Simple LLM-based Multi-Agent Framework for Software Evolution

• 🔍 Visualization of Multi-Agent Processing

All commands below have been tested on a Linux-64 machine. The specified package versions are recommended, as they have been verified to work.

conda create -n magis python=3.11 -y
conda activate magis
conda install redis==5.0.3 -y  # If you need to cache repeated requests to the same LLM
pip install openai==0.27.8 pygments==2.17.2 pydeps==1.12.17 tiktoken==0.5.1 redis==5.0.1 flask==3.0.0 datasets==2.16.1

To use MAGIS on a Git repository, please run the magis.py script.
You can view options and usage instructions by executing: python magis.py --help.

To reproduce our experiments on SWE-bench, please refer to SWE-bench.md.

If you find MAGIS useful for your research or applications, feel free to give us a star ⭐ or cite us using:

@inproceedings{magis,
  author       = {Wei Tao and
                  Yucheng Zhou and
                  Yanlin Wang and
                  Wenqiang Zhang and
                  Hongyu Zhang and
                  Yu Cheng},
  title        = {{MAGIS:} LLM-Based Multi-Agent Framework for GitHub Issue Resolution},
  booktitle    = {NeurIPS},
  year         = {2024}
}

• We thank the prior work, SWE-bench and their responses for my question.

• We gratefully acknowledge OpenMoji for providing the emojis used in our demonstrations.