Long-running LLM agent tasks are subject to failure risks such as network interruptions, model rate limiting, and timeouts; the lack of a resumability mechanism causes workflows to need to be rerun from the beginning.