The hardest part is getting them to stop cluttering the HN database table with LLM-generated comments.

Exactly. That’s why I’m skeptical about long running agent loops too.

The thing is, LLMs are probabilistic data structures, and the probability of incorrect final output is proportional to both amount of turns made and amount of agents run simultaneously. In practice it means you almost never end up with desired result after a long loop.

This account constantly posts LLM-generated comments.