Your next feed refresh might get a second opinion from an AI agent you never asked to look at it.
A new arXiv paper studies what the authors call personal-agent mediated recommendation. A platform's recommender ranks items using only what it can see on its own service. A personal AI agent, authorized by the user, then adjusts that ranking using the user's history from other platforms before the final list reaches the screen. The researchers built a benchmark called MediateRec, with both simulated cross-platform setups and a real test that keeps the platform and the agent from seeing each other's data. To train agents to make smarter edits, they also introduce a method called Personal Attribution Mediation Optimization, which hides a user's cross-platform history during training to measure whether an edit actually helped, then adjusts rewards without punishing correct platform calls.
This matters because the trade-off is not hypothetical. The study found that even strong proprietary language models sometimes override a platform's ranking when they should not, not just when they should. That is the quiet risk in letting an agent fix your feed: good platform data can get outvoted by an agent working from an incomplete picture.
Call it the next round in the fight over who controls your attention, except this time it is your own software doing the overriding.