Ask HN: Shifting Reoccuring Agentic Workflow to ML Process?

I know I can simply ask myaAgent or Google this question but would prefer to ask HN for any practical real examples used day to day.

Question: For the people who build agentic workflows, how do I implement a ML habit where the agent learns from itself, can iterate and improve on it's own? I would still oversee and validate but would love to learn ML implementation to agents. Am I just overthinking this and can write it in the prompt? Any help is appreicated!