The post explores a hybrid workflow‑agent pattern where a predefined sequence coexists with an autonomous agent. It begins by contrasting workflow and agent paradigms: workflows are deterministic, easy to control, and suited for known steps, while agents are flexible and adapt to intermediate findings. Many real apps need both. The case study builds an LLM application for hyperparameter tuning of classification models. The solution is split into three stages:

1. Prepare the Experiment – A single LLM call (using GPT-5.4) translates a plain‑language request and dataset summary into an experiment brief. The output is a structured Pydantic schema with objective, primary_metric, and cross‑validation folds.

2. Hyperparameter Tuning – An agent, also powered by GPT-5.4, dynamically selects classifier configurations, evaluates them with a tool that runs scikit-learn cross‑validation, and records scores. The agent’s output includes model name, hyperparameters, and rationale.

3. Summarization – A final LLM call turns the brief, trial history, and recommendation into a report.

By embedding an agent only at the exploratory stage, the pattern blends workflow clarity with agentic flexibility, enabling adaptive model experimentation while keeping the overall flow predictable.