Skip to content
Projects

bingreeky/MaAS has 292 GitHub stars and 29 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-19 07:10 UTC. Last push 2025-11-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

[ICML'25 Oral] Multi-agent Architecture Search via Agentic Supernet

Developer Tools

README.md

Original

MaAS: Multi-agent Architecture Search via Agentic Supernet

πŸ“° News

  • πŸŽ‰ Updates (2025-05-03) MaAS is accepted as ICML'25 Oral (Top ~1% among 12,107 submissions)!
  • 🚩 Updates (2025-02-06) Initial upload to arXiv (see PDF).

πŸ€” What is Agentic Supernet?

We for the first time shift the paradigm of automated multi-agent system design from seeking a (possibly non-existent) single optimal system to optimizing a probabilistic, continuous distribution of agentic architectures, termed the agentic supernet.

MaAS

πŸ‘‹πŸ» Method Overview

Building on this concept, we propose MaAS, which dynamically samples multi-agent systems that deliver satisfactory performance and token efficiency for user queries across different domains and varying levels of difficulty. Concretely, MaAS takes diverse and varying difficulty queries as input and leverages a controller to sample a subnetwork from the agentic supernet for each query, corresponding to a customized multi-agent system. After the sampled system executes the query, MaAS receives environment feedback and jointly optimizes the supernet’s parameterized distribution and agentic operators.

framework

πŸƒβ€β™‚οΈβ€βž‘οΈ Quick Start

πŸ“Š Datasets

Please download the GSM8K, HumanEval, MATHdatasets and place it in the maas\ext\maas\data folder. The file structure should be organized as follows:

data
└── gsm8k_train.jsonl
└── gsm8k_test.jsonl
└── ......

πŸ”‘ Add API keys

You can configure ~/.metagpt/config2.yaml according to the example.yaml. Or you can configure ~/config/config2.yaml.

llm:
  api_type: "openai" 
  model: "gpt-4o-mini" 
  base_url: ""
  api_key: ""

🐹 Run the code

The code below verifies the experimental results of the HumanEval dataset.

python -m examples.maas.optimize --dataset HumanEval --round 1 --sample 4 --exec_model_name "gpt-4o-mini"
python -m examples.maas.optimize --dataset HumanEval --round 1 --sample 4 --exec_model_name "gpt-4o-mini" --is_test True

πŸ“š Citation

If you find this repo useful, please consider citing our paper as follows:

@article{zhang2025agentic-supernet,
  title={Multi-agent Architecture Search via Agentic Supernet},
  author={Zhang, Guibin and Niu, Luyang and Fang, Junfeng and Wang, Kun and Bai, Lei and Wang, Xiang},
  journal={arXiv preprint arXiv:2502.04180},
  year={2025}
}

πŸ™ Acknowledgement

Special thanks to the following repositories for their invaluable code and prompt.

Our prompt is partially adapted from ADAS, AgentSquare, and AFLOW. Our code and operators are partially adapted from AFLOW.