A new library wants AI agents to stop guessing at machine-learning code and start querying it directly.
Researchers have released TuiML, an open-source machine-learning library built specifically for AI agents rather than human programmers. Unlike scikit-learn or Weka, every algorithm in TuiML describes itself through machine-readable metadata and parameter schemas, letting an agent search the library, inspect what a component does, and assemble a validated workflow instead of recalling APIs from memory. Every call gets validated, seeded, and logged, and full sessions export as runnable notebooks. A single specification layer feeds the Model Context Protocol, agent-framework adapters, a Python API, a command line, and local model serving, and the paper reports TuiML holds its own against scikit-learn and Weka on prediction benchmarks.
The real fix here isn't accuracy, it's plumbing. Agents writing code against libraries meant for humans hit errors only at runtime and forget what they tried between turns. Making the library itself machine-readable turns trial-and-error into something closer to a documented, reproducible experiment.
It's also part of a bigger pattern: rebuilding developer tools with agents, not people, as the primary user, the same logic behind MCP itself. Whether TuiML becomes the norm or a research curiosity depends on something the benchmarks can't measure: whether agent frameworks actually adopt a whole new library instead of just getting better at using the old ones.