Traditional recommender systems are usually hosted on platforms, where operators can boost engagement through clickbait, filter bubbles, or even fake news, often at the expense of genuine user needs. To counteract platform lock‑in, the notion of user‑side recommender systems has emerged: users run their own models and are no longer bound by platform interests. Building such systems is challenging, especially when personalization requires extra user data.
AgentRecommender proposes to leverage the investigative capability and internal knowledge of large language model (LLM) agents to construct user‑side recommenders without additional data. An LLM agent can, based on a natural‑language preference description supplied by the user, actively retrieve, reason, and generate a list of candidate items. Users only need to provide a preference prompt in their local environment; the system then invokes the LLM agent to perform content filtering, relevance scoring, and ranking, delivering recommendations that match personal interests.
The core workflow consists of:
- User writes a natural‑language preference description;
- LLM agent parses the description and searches its knowledge base for matching item features;
- The agent evaluates candidate items using its internal model;
- Evaluation results are returned to the user‑side system for final ordering.
All operations run locally or in a trusted compute environment, preserving user privacy and preventing platform interference in the recommendation logic.
Review: AgentRecommender demonstrates the feasibility of LLM agents in user‑side recommendation scenarios, offering a novel technical route toward truly user‑centric personalization.