Agentic programming workflows rather remind me of genetic programming. The agent has a validation step that looks like a fitness function, and both run iterative trials.
There's something pleasingly self-referential about configuring an agent by talking to it.
> Remember that [fact about data formats].
Agent: Noted.
... realise nothing happened ...
> Write to your rules file that when I say "remember", I want you to write to the rules file.
After further playing with my LLM project, I'm surprised how hard it is to tune with system prompts.
My agent kept saying "obviously" even though my prompt said "helpful, professional". Eventually I found that a "courteous" prompt gets "of course" which is better but not ideal.