A new research paper proposes AgentRecommender, a way to build a recommendation engine - powered by an AI agent - that works for you instead of a platform's ad business.
Recommender systems on platforms like social feeds and video sites are built to serve the platform, not the reader. That setup produces clickbait, filter bubbles, and the spread of fake news, because those things drive engagement even when they don't help any individual user. User-side recommenders - systems a person runs themselves instead of relying on a platform's algorithm - have been proposed before as a fix, but building one has always meant collecting fresh data about that specific person's tastes. AgentRecommender skips that step: it uses a large language model agent's ability to investigate content and its existing built-in knowledge to assemble a personalized recommender without any additional data collection, according to the researchers.
The interesting part isn't the engineering, it's the incentive shift. A recommender that answers to the reader instead of an advertiser is a different product, not just a retuned algorithm - the same instinct behind ad blockers and RSS readers, now applied to ranking instead of blocking. The catch is the one every user-side tool eventually hits: it only works as well as the data platforms are willing to expose to an outside agent.
This is a research proposal, not a shipping browser extension, and whether it goes anywhere depends less on the AI agent's cleverness than on whether platforms tolerate a tool built to route around their own ranking decisions.