Introduces TuiML, a self-contained ML library with native algorithms and agent-native interfaces that expose capabilities, surface errors early, and maintain experimental state.
TuiML is an open-source, agent-native machine learning library developed by the University of Waikato, designed to allow AI agents to autonomously discover, train, tune, and serve models without writing glue code. Unlike traditional libraries like scikit-learn, TuiML exposes 300+ components (algorithms, preprocessors, metrics) as typed Model Context Protocol (MCP) tools with machine-readable JSON schemas, enabling agents to validate parameters, discover algorithms by task or data shape, and maintain persistent experimental state.
Key features include structured error recovery, where agents receive specific suggestions for fixing invalid calls, and local-first execution that keeps data and models on the user's machine. Benchmarks indicate TuiML is predictively competitive with scikit-learn and Weka, offering comparable accuracy while providing a unified interface for classification, regression, clustering, time-series, and anomaly detection tasks.
TuiML: Machine Learning for AI Agents introduces a self-contained machine learning library designed around the needs of AI agents rather than only human developers. Its central premise is that agent-driven ML workflows require interfaces that are explicit, composable, and robust to autonomous use. The library provides native algorithms alongside agent-native interfaces that expose available capabilities, make errors visible early, and preserve experimental state across steps, allowing an agent to inspect what operations are possible, understand failures without hidden side effects, and continue or revise an experiment in a controlled way.
The key contribution is a design shift from traditional, human-centric ML APIs toward interfaces optimized for agentic experimentation. By treating capability discovery, error surfacing, and state management as first-class concerns, TuiML aims to reduce common failure modes in agent-based ML work, such as hallucinated API calls, ambiguous failure states, and loss of context during iterative experiments. This matters because it supports more reliable, reproducible, and composable pipelines in which agents can perform data inspection, model training, evaluation, and debugging with clearer feedback loops and less dependence on fragile prompt-driven conventions.