When the User Is Not Always Human: Recommender Systems in the Agentic Era

Abstract

For decades, recommender systems (RecSys) have been optimized to serve human users, leveraging behavioral data to predict preferences. This assumption is now under pressure. The rapid deployment of autonomous agents powered by large language models (LLMs) introduces a paradigm shift: recommendation consumers are expected to be increasingly a mixture of humans and authorized agents acting on their behalf as client-side proxies. This paper characterizes the resulting tripartite ecosystem spanning humans, agents, and platforms, and formalizes the distinct interaction types and optimization objectives that emerge across these relationships. We then turn a critical eye on the past decade of human-centric RecSys research, identifying foundational blind spots — in algorithm design, dataset construction, and evaluation practice — that must not be carried into the agentic era. Building on these lessons, we outline four research directions: task-centric recommendation, protocol-level evaluation, cross-platform brokerage, and human-agent preference elicitation. This shift is not merely a change in interface. It is a mandate to rebuild recommender systems from the ground up.