I tend to think of LLM agentic loops as a version of a genetic algorithm. The LLM acts as a mutator function, and the harness provides selection pressures. The LLM creates candidate samples which get tested and then results are fed back LLM to iterate on. And this loop will eventually converge on a solution that passes the acceptance criteria. This is precisely why LLMs work so unreasonably well for agentic coding now.