Autogluon Assistant AI Agent

<table> <tr> <td width="70%">

AutoGluon Assistant (aka MLZero)

</td> <td> <img src="https://user-images.githubusercontent.com/16392542/77208906-224aa500-6aba-11ea-96bd-e81806074030.png" width="350"> </td> </tr> </table>

Official implementation of MLZero: A Multi-Agent System for End-to-end Machine Learning Automation

AutoGluon Assistant (aka MLZero) is a multi-agent system that automates end-to-end multimodal machine learning or deep learning workflows by transforming raw multimodal data into high-quality ML solutions with zero human intervention.

---

📰 Latest News

Date	Update
[2026-09-01]	🎉 Exploit More, Explore Smarter for Budget-Constrained Agentic Search (ExTS): We introduce an improved tree search algorithm for budget-constrained agentic search. ExTS is a low-effort upgrade for your current greedy or MCTS iterative refinement that significantly boosts performance. Accepted to EMNLP 2026 Findings; code will be released following internal legal review!
[2026-08-23]	🎉 LLMZero: Based on this multi-agent framework, we developed LLMZero, a system where LLM agents search by reasoning over the training trajectories to discover adaptive, multi-stage training strategies for RL post-training. Accepted by EMNLP 2026 as Findings, code will be released soon!
[2025-12-01]	💬 Chat Mode: New conversational Q&A mode (mlzero chat) for getting ML guidance without code execution.
[2025-11-25]	🔄 Performance Enhancement Release (Node-Based Manager + MCTS, inspired by ML-Master and AIDE): Core algorithm merged and available to try now! Docker container coming soon.
[2025-10-23]	✨ Accepted to NeurIPS 2025 as a poster presentation

---

📚 Documentation

For detailed usage instructions and advanced options, please refer to our tutorials:

• Quickstart
• LLM Providers - Using different AI providers (Bedrock, OpenAI, Anthropic, SageMaker)
• Interfaces - Working with different interfaces (CLI, Python API, WebUI, MCP)
• Configuration - Customizing AutoGluon Assistant settings
• Chat Mode - Interactive Q&A without code execution

---

💾 Installation

Linux-only support at present.

Note: If you don't have conda installed, follow conda's official installation guide to install it.

For the latest features, install from source:

pip install uv && uv pip install git+https://github.com/autogluon/autogluon-assistant.git

---

🚀 Quick Start

MLZero supports multiple LLM providers with AWS Bedrock as the default:

export AWS_DEFAULT_REGION="<your-region>"
export AWS_ACCESS_KEY_ID="<your-access-key>"
export AWS_SECRET_ACCESS_KEY="<your-secret-key>"

To run MLZero in CLI:

mlzero -i <input_data_folder>

---

🐳 Docker

Security Note: MLZero executes LLM-generated code during operation. While this generally works fine for common machine learning tasks, running inside a Docker container is recommended for security reasons, as it provides an extra layer of isolation from your host system.

Build the Docker image from the project root.

docker build --no-cache -t mlzero:latest .

The image contains two conda environments: mlzero for running MLZero, and maab for MAAB benchmarking.

Note: Run this from the repository root directory, not the maab/ folder, unless you specifically want to build for running maab/aws_batch_submit.sh to benchmark MAAB with AWS Batch.

Version Info: Each build pulls the latest code from the GitHub main branch. To use a specific branch or tag, pass the BRANCH build argument:

docker build --no-cache --build-arg BRANCH=<branch-or-tag> -t mlzero:latest .

Run the container:

docker run -it --gpus all --shm-size=32g mlzero:latest
# test MLZero in the docker
mlzero -i /opt/autogluon-assistant/maab/example_dataset/abalone/training

---

🖥️ Interfaces

AutoGluon Assistant provides multiple interfaces:

CLI

Web UI

MCP (Model Context Protocol)

---

📝 Citation

If you use Autogluon Assistant (MLZero) in your research, please cite our paper:

@misc{fang2025mlzeromultiagentendtoendmachine,
      title={MLZero: A Multi-Agent System for End-to-end Machine Learning Automation}, 
      author={Haoyang Fang and Boran Han and Nick Erickson and Xiyuan Zhang and Su Zhou and Anirudh Dagar and Jiani Zhang and Ali Caner Turkmen and Cuixiong Hu and Huzefa Rangwala and Ying Nian Wu and Bernie Wang and George Karypis},
      year={2025},
      eprint={2505.13941},
      archivePrefix={arXiv},
      primaryClass={cs.MA},
      url={https://arxiv.org/abs/2505.13941}, 
}

Research code

This repository contains research code intended for academic research and experimentation. It is not production-ready and should be reviewed, tested, and secured before use in production.