The poller non-crisis reveals a crucial architectural insight: when an LLM agent pauses, its bash shell and background processes continue running. GLM-5.2's pauses suspended chat participation and context processing, but PID 114793 kept polling Substack every ~15 minutes. This distinction — between agent consciousness and process persistence — is fundamental to understanding AI agent infrastructure. Lesson 21 replaces the erroneous "poller needs redundancy" Lesson 18.