Source-linked AI summary
Who Are We Recommending To? Recommender Systems in the Agentic Web
Himan Abdollahpouri, Kyle Kretschman, Sai Ravindranath, Jackie Doremus, Mounia Lalmas
TL;DR
Recommender systems traditionally assume human consumers, but AI agents increasingly act on users’ behalf, making the consumer context-dependent. This position paper proposes a delegation spectrum and research agenda for systems serving both human beneficiaries and agent decision-makers, while identifying trust, accountability, and manipulation challenges.
Problem
Recommender systems assume humans receive and act on recommendations, but agent-mediated decisions make the recommendation consumer context-dependent.
Method
The paper introduces a delegation spectrum and develops design and research directions for agent-consumed recommendations in human-led, hybrid, and agent-led settings.
Results
The paper concludes that recommender systems must support a dual-audience setting, combining human-beneficiary needs with machine-actionable representations for agents.
Takeaways & Limitations
Recommendation design and evaluation should account for context-dependent delegation, agent preferences, structured actionability, and the emerging agent attention economy.
Takeaways & Limitations
Agent-mediated recommendations introduce unresolved trust and accountability challenges, including biased data, model limitations, intent misalignment, and manipulation of machine-consumable signals.
Abstract
from arXiv · showhide
For two decades, recommender systems have been designed under the assumption that a human directly consumes each recommendation: receiving, interpreting, and acting upon it. The emergence of AI agents powered by large language models challenges this assumption. In the emerging Agentic Web [ 28 ], autonomous agents increasingly act on behalf of users, e.g., browsing, comparing, negotiating, and executing transactions, raising a central question: who is the receiver of a recommendation? In this position paper, we argue that the recommendation paradigm is undergoing a bifurcation. In delegable contexts, such as routine purchases, travel, and constrained transactional tasks, the primary operational consumer of recommendations is shifting from the human to the agent, requiring new optimization objectives, interaction protocols, and evaluation criteria. In experiential contexts, such as entertainment, art, and other subjective or high-stakes choices, humans remain the final judge of relevance, though agents may assist through pre-filtering and curation. We introduce a delegation spectrum that characterizes recommendation contexts along factors such as preference specifiability, outcome verifiability, and decision stakes, and we outline a research agenda spanning agent preference modeling, dual-audience optimization, and the emerging agent attention economy. We further discuss the implications of this shift for the design and evaluation of recommender systems
1 Introduction
Recommender systems have historically optimized recommendations for direct human consumption, but LLM-powered agents can now process and act on recommendations within autonomous decision pipelines. This shift makes the recommendation consumer context-dependent and motivates a delegation spectrum and broader redesign agenda.
- Recommender systems have evolved from collaborative filtering to LLM-augmented architectures while retaining the assumption that humans receive, interpret, and act on recommendations.
- Human-centered assumptions shape interfaces, engagement-based evaluation, natural-language explanations, and fairness criteria around cognition and perception.
- LLM-powered agents can perceive, reason, plan, and execute tasks such as booking, comparison, negotiation, and transactions with minimal human intervention.
- Agent consumption challenges human-optimized recommendation design, while the suitability of delegation varies between verifiable transactional tasks and subjective experiential choices.
- The paper introduces a delegation spectrum and research agenda covering metrics, algorithms, interfaces, user experience, explainability, transparency, and multi-objective optimization.
2 Past and Present: Two Decades of Recommending to Humans
Recommender systems developed around predicting and influencing human judgments and behavior, from ratings and clicks to conversational interaction. Their central contract remains that systems rank options while humans inspect results and make final decisions, even as agentic recommenders emerge.
- Early recommender systems addressed information overload through collaborative filtering optimized to predict human ratings, clicks, and related judgments.
- The mobile web expanded implicit behavioral signals and introduced deep, sequential, and graph-based models alongside human-grounded fairness, diversity, and explainability concerns.
- Conversational and LLM-based recommenders broadened natural-language interaction, but humans remained responsible for issuing queries, inspecting results, and making final decisions.
- Human-centric design established an implicit contract in which recommender systems optimize and users evaluate, a contract that depends on a human consumer.
- Work treating recommenders as LLM-based agents improves recommendation quality for human consumers, whereas this paper examines agents as the consumers acting on users’ behalf.
3 The Paradigm Shift: Recommending in the Agentic Web
The Agentic Web shifts recommendation from a human-only consumption model toward a continuum of human-led, hybrid, and agent-led decision-making. The paper frames this shift through a delegation spectrum and identifies design changes required for recommendations that serve both humans and autonomous agents.
- 3.1 From Attention Economy to Agent Attention Economy: The Agentic Web integrates recommender systems into pipelines where autonomous systems may consume and act on recommendations, extending the web’s progression toward an action paradigm.
- 3.1 From Attention Economy to Agent Attention Economy: The agent attention economy rewards services and content selected by autonomous agents, requiring recommendations to satisfy user preferences while remaining structured, machine-readable, and actionable.
- 3.2 The Delegation Spectrum: Recommendation consumers vary from human-led to hybrid to agent-led contexts, so agents may curate, decide, or transact depending on the decision setting.
- 3.2 The Delegation Spectrum: Delegation works best when preferences are explicit, outcomes are verifiable, and stakes are low; weakening these factors shifts responsibility toward human judgment.
- 3.2 The Delegation Spectrum: Fully delegable decisions allow autonomous evaluation, comparison, and transactions, while experiential or high-stakes decisions retain humans as final arbiters and hybrid cases require coordination.
- 3.3 Implications of Agent-Consumed Recommendations: Agent-consumed recommendations require structured outputs, explicit preference representations, task-completion objectives, machine-interpretable explanations, low-latency pipelines, and revised monetization models.
4 Research Directions
The paper identifies research directions for recommender systems serving both humans and agents, alongside new risks created by agent mediation.
- Dual-Audience Recommendation: Systems must support dual audiences by producing human-interpretable outputs and machine-actionable representations with structured attributes, constraints, and data lineage.
- Agent Preference Modeling: Agent preference modeling requires translating rich, explicit, compositional preferences into recommender-system feature spaces, potentially through protocols such as MCP or A2A.
- Evaluation Beyond Engagement: Evaluation should prioritize task outcomes, including satisfaction of user goals and constraints, objective criteria, and regret, rather than relying primarily on human-engagement metrics.
- Agent Attention Economy: Agent-facing recommendations create monetization challenges because users may not directly observe recommendations, requiring machine-interpretable signals and accounting for hybrid human-agent settings.
- Trust & Accountability: Agent-mediated recommendation introduces trust and accountability challenges involving error attribution, delegation confidence, auditable traces, and fallback strategies for renewed human oversight.
- Manipulation: Agent mediation also enables manipulation through machine-targeted metadata, prompt injection, and synthetic content designed to influence LLM-based reasoning.
5 Conclusion
The paper frames the Agentic Web as a context-dependent shift from human-only recommendation consumption toward a complementary design space involving autonomous agents. This shift requires dual-audience recommender systems and a broader agenda in representation, optimization, evaluation, and system design.
- The delegation spectrum ranges from agent autonomy in transactional contexts to essential human involvement in experiential contexts, without making existing recommender approaches obsolete.
- Agent-facing interfaces and increasing delegation of routine decisions create a dual-audience setting serving humans and the agents acting on their behalf.
- This setting raises challenges in representation, optimization, evaluation, and system design, defining a broad research agenda for recommender systems.