AutoGluon Assistant: No-Code AutoML Built on Multi-Agent Collaboration

Traditional AutoML automated model selection and hyperparameter tuning, but users still had to write code and clean data themselves. AutoGluon Assistant aims to remove this final barrier by building on a multi-agent system called MLZero to construct training pipelines for tabular, image, text, and time series data using only natural language descriptions.

This system ranked 10th as a fully automated framework in the 2024 Kaggle AutoML Grand Prix, becoming the only automated agent to record a score in the competition. It also showed strong performance in both solution quality and success rate, achieving a 92% success rate on the Multimodal AutoML Agent Benchmark and an 86% success rate on the external benchmark MLE-bench Lite.

At its core is a structure where four modules divide the work:

• perception: Interprets file structures and content to identify data types and tasks. Even when the target variable isn't clear, as in CSV files, it analyzes column distributions and meaning to infer problem types such as classification, regression, or segmentation.
• semantic memory: Stores AutoGluon's features, API patterns, and best practices in structured form. Based on this, it selects the appropriate tools for tasks like semantic segmentation.
• episodic memory: Accumulates execution history and success/failure cases to provide debugging context. When errors occur, it references previous attempts and results to determine the direction for the next fix.
• iterative coding: Executes generated code and iteratively improves it by reflecting results or errors. It maintains a feedback loop until success or until the maximum number of iterations is reached.

Using an example of a medical X-ray and segmentation mask, the authors explain that even if a user simply types "find the lesion area," perception understands the task, semantic memory selects an appropriate MultiModalPredictor, and iterative coding fixes mask format mismatches to complete training. In this process, users don't need to write code, but they can input domain knowledge at each iteration step for correction if needed.

Evaluation proceeded along two axes. MLE-bench Lite is an external standard benchmark comprising 21 tasks from previous Kaggle competitions; AutoGluon Assistant successfully submitted 18 out of 21, recording the highest success rate of 86%, and also ranked first in overall solution quality with an average rank of 1.43. Its own benchmark, the Multimodal AutoML Agent Benchmark, consists of 25 tasks featuring more raw data, higher noise, format inconsistencies, and even multilingual, multi-table, and large-scale data, on which it also achieved a 92% success rate.

Performance was maintained even in resource-constrained environments. Implemented with an 8B-class small LLM, it recorded a 45.3% success rate, demonstrating greater efficiency than many agents using larger models.

The approach is designed to easily fit into existing workflows, supporting CLI, Python API, Web UI, and MCP. It has been released as open source, and the technical details are documented in a NeurIPS 2025 paper.