I'm doing these days at work, I feel like they are sometimes surprised when I say 'building CatBoost classifiers' or 'fitting isolation forests'. But this is what I spend a good chunk of my time doing. My company's product is a sophisticated agentic AI platform, but I don't do much prompt engineering. A year ago, I'd be concerned about classical ML's future, but agentic AI needs classical ML more than we thought.

An AI Agent combines LLMs and tooling for workflows with minimal human intervention. The LLM interfaces between users and tools, enabling functionality beyond token generation. Key is tooling — agents need tools like data retrieval, RAG knowledge bases, and query construction. Classical ML models are also valuable tools.

Classical ML offers advantages: Accuracy — models are more accurate than LLMs guessing. Interpretability — decisions can be identified and validated. Cost — classifiers and regression models are lightweight and cheap. Precision — you control training and tuning. Control — no third-party data or infrastructure risk.

Two architectural choices: Direct Calls — agent calls model directly with structured output. Database Access — pre-calculate inferences as scheduled jobs for agent retrieval. Both require clear documentation of model purpose and capabilities.