Who Are We Recommending To? Recommender Systems in the Agentic Web

TL;DR: Recommender systems have always assumed a human recipient butAI agents break that assumption. In our RecSys 2026 paper, we introduce a delegation spectrum from human-led to agent-led recommendation. Where a task falls on this spectrum depends on how specifiable preferences are, how verifiable outcomes are, and how reversible the decision is. This raises open questions about how recommendations should be represented, evaluated, and governed when the immediate consumer is an agent acting on a user’s behalf.

The “Human” Assumption

For nearly three decades, recommender systems have evolved from collaborative filtering and matrix factorization to deep learning, sequential models, and now  systems augmented by large language models. Yet throughout this evolution, one assumption has remained largely unchanged: the recipient of a recommendation is a human.

This assumption has shaped the entire recommender systems stack. Interfaces are optimized for human inspection through ranked lists and feeds. Evaluation relies on engagement proxies such as clicks, dwell time, and conversions. Explanations are designed for interpretability through natural language. Fairness is defined in terms of exposure, diversity, and user experience. In short, recommender systems have been designed around human cognition and perception. AI agents are beginning to challenge this assumption.

Agents as New Consumers

Consider a user who tells an AI agent: “Put together a 2-hour dinner party playlist,  jazzy and downtempo, mostly tracks I haven’t heard before, but sprinkle in some of my favorites so it doesn’t feel unfamiliar.” The agent might receive candidates from discovery, personalization, and mood-based services, evaluate them against the user’s constraints (e.g., genre, novelty, familiarity balance, duration) and assemble the playlist without requiring the user to browse through hundreds of tracks. Or consider asking a travel agent: “Find the cheapest nonstop or max 1-stop flight to Minnesota in October and book a refundable hotel near downtown.” The agent receives recommendations from travel services, evaluates the alternatives against the user’s preferences and constraints, and makes the booking.

In both cases, the human is the ultimate beneficiary, but the immediate consumer of the recommendations is an AI agent.

Agent-consumed recommendations can also emerge inside recommendation architectures themselves. Several specialized services, each generating candidates for a different purpose,  might send their outputs not to a user interface, but to an LLM-based agent that reasons across them, applies contextual constraints, and decides what to present to the user. Here, the agent mediates rather than replaces human judgment: filtering, resolving trade-offs, and curating a set for the user to evaluate. This hybrid pattern of “not fully human-led, not fully agent-led” is where much of the near-term design space lies.

These scenarios raise a fundamental question: who are we recommending to?

In our paper, “Who Are We Recommending To? Recommender Systems in the Agentic Web”, accepted at the 20th ACM Conference on Recommender Systems (RecSys 2026) in the Past Present and Future track, we explore how recommender systems may need to evolve when recommendations are consumed not only by humans, but also by AI agents acting on their behalf. As a company where recommendation is central to the experience we deliver to the listeners, these questions are directly relevant to how we think about the future of our systems.

It is worth noting that this question is distinct from recent work on the recommender system itself as an LLM-based agent, where agentic capabilities such as planning, memory, and multimodal reasoning are used to improve recommendation quality for human consumers. Our focus is orthogonal: we ask what changes when the consumer of the recommendation is an agent.

The Delegation Spectrum

When we began examining this question, we initially considered whether agents might eventually render human-facing recommendation interfaces unnecessary. We have concluded that this is unlikely. The degree to which recommendations can be delegated to agents varies considerably across decision contexts.

Consider the difference between “Reorder my usual printer paper” and “Find a gift for my sister’s birthday.” In the first case, preferences are explicit, the outcome is easy to verify, and the cost of a mistake is small. In the second, what makes a recommendation good depends on subjective judgment, personal context, and experiential factors that resist full specification.

We formalize this variation through a delegation spectrum, ranging from human-led to hybrid to agent-led recommendation. Where a particular context falls on this spectrum depends on three factors:

• Preference specifiability: How precisely can the user’s preferences and goals be articulated in advance? “Under $500, nonstop, arriving before 6pm” is explicit. “A dress that feels right for my friend’s wedding” is constructed through exposure and reflection.

• Outcome verifiability: Can the quality of a recommendation be assessed against objective, machine-checkable criteria, or does it require subjective human experience? An agent can verify that a battery meets a voltage specification. It cannot verify that a novel will move the reader.

• Stakes and reversibility: What are the consequences of a wrong decision, and how easily can it be reversed? A $6 pack of batteries or a streaming recommendation tolerates agent error. A medical decision or a home purchase demands human oversight even when preferences and outcomes are relatively well-specified.

Figure 1: The delegation Spectrum in recommender systems in agentic web

Taken together, these three factors yield a simple heuristic: delegation works best when preferences are explicit, outcomes are verifiable, and stakes are low. As any of these factors weakens, the appropriate role of the agent shifts toward filtering and assisting, while the human retains more of the final judgment.

This distinction matters most for experiential domains, where agents may narrow or contextualize recommendations, but cannot substitute for human judgment about whether a song resonates, a film is enjoyable, or a gift feels meaningful.

Designing for Agent-Consumed Recommendations

If recommendations increasingly serve as inputs to AI agents, several parts of the recommender systems stack may need to evolve. In the paper, we identify a number of these, highlighted here.

Dual-audience recommendation. When a recommendation benefits a person but is first processed by an agent, human-facing outputs (e.g., ranked lists, cards, natural-language explanations) may need to coexist with machine-actionable representations containing structured attributes, constraint annotations, and provenance metadata.

Richer preference representations. Recommender systems have traditionally inferred preferences from behavioral signals such as clicks, streams, and purchases. Agents can maintain explicit, compositional representations of a user’s goals, constraints, and context,  potentially communicated through emerging protocols such as MCP or A2A. How recommender systems should ingest and leverage these richer representations is an open question.

Evaluation beyond engagement. Metrics such as precision, recall, NDCG, and click-through rate reflect a world in which humans directly consume recommendations. In agent-mediated settings, the relevant question shifts from engagement to goal achievement: did the recommendation enable the agent to satisfy the user’s goals and constraints? In hybrid settings, evaluation must also capture interaction dynamics such as override rates and intervention frequency.

Trust and accountability. Failures can arise at multiple points when an agent reasons over a recommendation, combines it with other information, and takes an action. Understanding whether a poor outcome was caused by the recommendation, the agent’s interpretation, or some other component creates challenges for transparency, auditing, and user control.

Monetization. Recommendation-driven advertising has historically relied on human attention. In agent-facing settings, where a human may never see the recommendation slate, these mechanisms lose effectiveness. This raises questions about how to distinguish intrinsic quality from paid influence in machine-readable representations.

Manipulation. Just as content providers learned to optimize for search engine algorithms, we should expect analogous optimization targeting AI agents, for example through inflated metadata, adversarial prompt injection, or synthetic content designed to exploit LLM-based reasoning.

From the attention economy to the agent attention economy

These changes point toward a broader structural shift. The web’s recommendation era has largely operated within an attention economy, where platforms and content providers compete for human attention. In an agent-mediated ecosystem, a parallel form of competition emerges: an agent attention economy, where services, tools, and content providers compete to be selected and invoked by autonomous agents rather than clicked by humans.

This shifts what it means for a recommendation to be effective. A recommendation optimized to attract a person’s attention (e.g., through visual salience, emotional appeal, or social proof)  may not be the same recommendation that an agent selects when reasoning over explicit constraints and objectives. 

Importantly, these two economies are unlikely to be entirely separate. Humans will directly consume recommendations in some contexts, agents will consume them in others, and many systems will operate somewhere in between. Recommender systems must increasingly support this dual-audience reality.

Looking ahead

For decades, recommender systems have operated under a consistent model: the system recommends, the human evaluates and acts. AI agents introduce new possibilities along this pipeline: as autonomous decision-makers for routine tasks, as intermediaries that filter and synthesize before recommendations reach users, or as reasoning components within recommendation architectures themselves. The result is not a simple transition from recommending to humans to recommending to agents, but a broader design space in which the consumer of a recommendation varies with the context.

For more information, read our paper:

Who Are We Recommending To? Recommender Systems in the Agentic Web

Himan Abdollahpouri, Kyle Kretschman, Sai Srivatsa Ravindranath, Jackie Doremus, Mounia Lalmas

RecSys 2026.