I think this is the natural endpoint - local models doing something like what Atlas and Codex are doing, acting like a firewall between the user and web. You don't need to wade through the crap online yourself, the AI extracts the useful signal for you, acts like a memory layer for your values and preferences. Don't like the feed ranking - use your agent to extract, filter and rerank by your criteria. Not a big fan of dark UI patterns? Not a problem anymore, the UI can be regenerated. Need a stronger model? Sure, your agent can delegate.
I see this as a big unbundling, since your agent has your ear now, not Google, not social networks, they lose their entry point status and don't control by ranking, filtering and UI what I see or what I can do. They can spread out searches to specialized engines, replace Google for search, walk above all social networks and centralize your activities so you don't have to follow each one individually. A wrapper or cocoon for the user, taking the ad-block and anti-virus role, protecting your privacy and carefully reducing your exposure to information leaks.
All of this only works if you can host your model. But this is where the trend is going, we can already see decent small models, maybe before 2030 we will be running powerful local models on efficient local chips.