Why do AI agents need to scrape so often, vs. aggressively caching or using archive.org or their own crawls of the internet?