Put the Agent Inside the Workflow

A hybrid pattern for building LLM applications combines the predictability of predefined workflows with the flexibility of autonomous agents. This approach is demonstrated through a case study on hyperparameter tuning for a classification problem. The process is structured into three stages: a workflow stage to prepare the experiment, an agentic stage to explore model configurations, and a final workflow stage to summarize the results. The agent is given tools to run experiments and can dynamically decide which configurations to test based on previous outcomes, providing control over the overall process while allowing for adaptive behavior where needed.