AgentRec: Multi-Agent Recommender Systems
- AgentRec is a research domain that leverages context-sensitive, similarity-based embeddings to select specialized LLM agents for varied tasks.
- It spans both agent routing for diverse tools and collaborative multi-agent conversational recommendation architectures.
- Empirical evaluations show improved retrieval accuracy and conversation efficiency while highlighting challenges in reproducibility and system design.
AgentRec denotes a cluster of research problems at the intersection of multi-agent systems and recommendation. In one explicit usage, it is an architecture for recommending which specialized LLM agent should handle a natural-language prompt by extending Sentence-BERT and retrieving over agent-specific prompt corpora (Park et al., 23 Jan 2025). In another, it is a next-generation LLM-powered multi-agent collaborative recommendation framework for conversational recommendation, built around specialized agents for conversation understanding, preference modeling, context awareness, and dynamic ranking (Ma et al., 2 Oct 2025). Related work extends the same selection logic to benchmarked agentic routing, enterprise architecture choice, and graph-based retrieval of tools and agents, which suggests that AgentRec is best understood not as a single algorithmic family but as a broader problem of context-sensitive agent selection under task, interaction, and governance constraints (Nizar et al., 22 Nov 2025).
1. Terminological scope and problem family
In the literature represented here, the term “AgentRec” appears in multiple senses. One sense concerns agent recommendation in the strict routing sense: given a natural-language prompt, select the most appropriate specialized LLM agent. A second sense concerns multi-agent recommendation systems: a team of agents collaborates to understand a user, track preferences, model context, and produce item recommendations. A third, broader sense appears in adjacent work where “AgentRec-style” reasoning refers to recommending agents, tools, or agent architectures for a task, even when the paper itself uses a different title (Park et al., 23 Jan 2025).
| Usage | Core object selected | Representative source |
|---|---|---|
| Agent routing | Specialized LLM agent | (Park et al., 23 Jan 2025) |
| Conversational recommendation | Recommended items via multi-agent collaboration | (Ma et al., 2 Oct 2025) |
| Agent/tool retrieval or architecture choice | Agents, tools, or agentic configurations | (Nizar et al., 22 Nov 2025) |
This semantic spread is reinforced by benchmark and systems papers. “AgentRecBench” is not a single recommender model but a benchmark ecosystem for “LLM agent-based personalized recommender systems,” while “AgentArch” treats the practical question of which agent architecture should be chosen for enterprise workflows as an empirical recommendation problem over orchestration, prompting, memory, and thinking-tool choices (Shang et al., 26 May 2025, Bogavelli et al., 13 Sep 2025). The term therefore spans both recommendation by agents and recommendation of agents.
2. Agent recommendation as routing among agents
The routing version of AgentRec starts from a multi-agent setting in which different agents expose different tools, workflows, reasoning strategies, or domain expertise. The central problem is to decide dynamically which agent should receive a task, rather than invoking agents in a fixed order or training a fixed closed-set classifier. “AgentRec: Agent Recommendation Using Sentence Embeddings Aligned to Human Feedback” frames this as an embedding-space problem: prompts appropriate for the same agent should lie near one another, and recommendation should be done by similarity to agent-specific prompt corpora rather than by a conventional output head (Park et al., 23 Jan 2025).
A related systems-level formulation appears in “Agent-as-a-Graph,” which argues that routing should not operate only at the coarse agent-description level when each agent may hide many MCP servers and tools. It represents agents and tools as a bipartite graph , retrieves both tool nodes and agent nodes, reranks them with type-specific weighted reciprocal rank fusion, and then traverses tool-to-agent ownership edges to return executable parent agents. On LiveMCPBenchmark, this raises Recall@5 to $0.85$ and nDCG@5 to $0.47$, with reported improvements of and over prior state-of-the-art retrievers (Nizar et al., 22 Nov 2025).
Enterprise work broadens the same routing question from “which agent should answer?” to “which architecture should be used at all?” “AgentArch” evaluates 18 architectural configurations across 6 models on 2 enterprise workflows and concludes that there is no one-size-fits-all agent architecture. The best acceptable pass@1 reaches on the simpler task and on the more complex task, with maximum pass only $0.0634$, indicating that architecture recommendation itself is an unresolved control problem rather than a solved systems default (Bogavelli et al., 13 Sep 2025). This supports a general interpretation of AgentRec as a family of context-dependent selection problems over agents, tools, or orchestrations.
3. “AgentRec” as SBERT-based agent recommendation
The paper “AgentRec: Agent Recommendation Using Sentence Embeddings Aligned to Human Feedback” proposes an embedding-based recommendation architecture built on SBERT, specifically all-mpnet-base-v2, with 768-dimensional embeddings (Park et al., 23 Jan 2025). The incoming prompt is optionally normalized through a “rephrase and respond” step into a concise task-focused sentence, then embedded and compared against cached corpora of prompts associated with each candidate agent. Each agent has a corpus
$0.85$0
and the query embedding $0.85$1 is scored against every $0.85$2 using cosine similarity: $0.85$3
The paper reports that single-neighbor matching is inadequate: using only the maximum cosine similarity in a corpus yields $0.85$4 top-1 accuracy, whereas arithmetic mean gives $0.85$5, geometric mean gives $0.85$6, and the best result comes from a generalized $0.85$7-mean with $0.85$8, reaching $0.85$9 top-1 accuracy (Park et al., 23 Jan 2025). The generalized $0.47$0-mean is
$0.47$1
The text states that a logarithmic generalized $0.47$2-mean is used for numerical stability, but the printed score equation is corrupted in the supplied text.
Training proceeds in two stages. First, the SBERT encoder is fine-tuned with triplets using BatchAllTripletLoss, so prompts assigned to the same agent cluster while prompts for different agents separate. Second, the paper introduces an RLHF stage intended to align the embedding space with human judgments about nuanced routing decisions. The RLHF implementation, however, is underspecified: the paper does not provide the exact reward model form, the RL algorithm, or the optimization objective. This omission is significant because the paper also reports that a supervised-only model at learning rate $0.47$3 reaches $0.47$4 top-1 accuracy—higher than the final AgentRec result—but behaves worse on edge cases such as distinguishing health, therapy, and fitness prompts (Park et al., 23 Jan 2025).
The dataset is synthetic and contains 8 agents with 1,250 prompts each, for a total of 10,000 prompts: tech support, cooking, math, gaming, therapy, reading, health, and fitness. It is split into 8,000 training and 2,000 testing examples, with the training split further partitioned into 6,000 for encoder fine-tuning and 2,000 for the reward model. Generation uses Llama-3.1-8B-Instruct with top-$0.47$5 sampling ($0.47$6), nucleus sampling ($0.47$7), repetition penalty $0.47$8, and temperature $0.47$9, followed by MinHash deduplication (Park et al., 23 Jan 2025).
Inference is computationally light because agent corpora are pre-embedded and cached. The paper reports less than 0 milliseconds per classification on test data, and about 1 milliseconds on average after CPU warm-up without the rewrite step (Park et al., 23 Jan 2025). The resulting architecture is not a fixed softmax classifier but a nearest-neighbor-style recommender over agent corpora, which is why the paper emphasizes adaptability to new classes, interpretability through nearest prompts, and computational cheapness.
4. “AgentRec” as multi-agent conversational recommendation
A distinct use of the name appears in “AgentRec: Next-Generation LLM-Powered Multi-Agent Collaborative Recommendation with Adaptive Intelligence” (Ma et al., 2 Oct 2025). Here the problem is not agent routing among specialized LLM workers, but interactive conversational recommendation under dynamic user preferences, conversation coherence, multiple ranking objectives, and real-time adaptation. The framework is a hierarchical multi-agent architecture with four specialized agents: Conversation Understanding Agent, Preference Modeling Agent, Context Awareness Agent, and Dynamic Ranking Agent, coordinated by an Adaptive Coordinator.
The model is organized into three phases: Parallel LLM-Agent Processing, Adaptive Intelligence Coordination, and Collaborative Ranking. The conversation state is represented by
2
the preference state by
3
the context state by
4
and the ranking output by
5
The coordinator computes dynamic agent weights using
6
and fuses per-agent scores as
7
The paper notes OCR corruption in several printed formulas, but these variable meanings are stated explicitly in the surrounding text (Ma et al., 2 Oct 2025).
A central systems contribution is the three-tier learning strategy. Tier 1, the Rapid Response Layer, handles simple queries with cached patterns and lightweight models, processes 8 of requests, and operates at sub-second latency. Tier 2, the Intelligent Reasoning Layer, handles 9 of queries with specialized agents at 0–1 seconds. Tier 3, the Deep Collaboration Layer, handles the remaining 2 with full multi-agent collaboration at 3–4 seconds (Ma et al., 2 Oct 2025). Routing is based on complexity scores computed from conversation history, user profile completeness, and query ambiguity, but the paper does not specify a formula, thresholds, or a gating architecture.
Empirically, the paper evaluates on DuRecDial, DuRecDial 2.0, and MultiWOZ. Reported table values are 5 Success@10, 6 Recall@10, 7 NDCG@10, and 8 Avg. Turns for AgentRec, compared with 9, 0, 1, and 2 for UniMIND. The abstract summarizes this as a 3 enhancement in conversation success rate, a 4 improvement in NDCG@10, and 5 better conversation efficiency at comparable computational cost, though the paper does not reconcile the slight discrepancy between the 6 abstract claim and the 7 Avg. Turns table row (Ma et al., 2 Oct 2025).
Despite the stronger empirical section relative to the SBERT-based AgentRec paper, major details remain absent. The paper does not specify the backbone LLMs, prompts, optimizer, learning rate, batch size, loss functions, online adaptation rule, or ablations. Its contribution is therefore strongest as an architectural proposal and benchmark result rather than as a fully reproducible optimization recipe (Ma et al., 2 Oct 2025).
5. Benchmarks, substrates, and adjacent research
Beyond papers explicitly titled AgentRec, a broader research ecosystem defines the surrounding problem space. “AgentRecBench” introduces an interactive textual recommendation simulator with a unified User-Review-Item network, three scenarios—classic, evolving-interest, and cold-start—and a comparison of 10 methods. It frames agentic recommenders as LLM-based agents operating in an environment 8 with an action space 9, thereby treating recommendation as a sequential decision problem rather than a static scorer (Shang et al., 26 May 2025). This benchmark infrastructure is important because it evaluates planning, reasoning, tool use, and memory as part of recommendation behavior, not only top-0 ranking.
“AgenticTagger” addresses a different but adjacent layer: item representation. It uses an architect–annotator multi-agent reflection loop to build a hierarchical, low-cardinality vocabulary of descriptors and then assigns those descriptors to items. The objective is to improve downstream recommendation through structured item representations, and the paper reports gains in generative retrieval, ranking, term-based retrieval, and critique-based recommendation (Xie et al., 5 Feb 2026). This suggests that AgentRec is not only about routing or coordination; it also depends on the quality of semantic representations on which routing or ranking operates.
Another adjacent line of work relocates the agent from the platform side to the user side. “iAgent” proposes a user-agent-platform paradigm in which an LLM agent acts as a protective shield between user and recommender system, enabling indirect exposure. The recommendation target is formalized as
1
with user instruction 2, external environment 3, and user-specific parameters 4. The stronger variant, Instruct5Agent, adds dynamic memory and improves both ranking metrics and protection-oriented measures such as ad filtering and popularity-adjusted ranking (Xu et al., 20 Feb 2025). Relative to AgentRec in the routing sense, this work treats the agent as an individualized mediator rather than as a backend selector.
In decentralized and tool-rich ecosystems, recommendation also depends on trust and context-aware competence attribution. “AgentReputation” argues that autonomous, strategic agents operating across heterogeneous software tasks require context-conditioned reputation cards, explicit verification regimes, and a policy engine for allocation, access control, and verification escalation. The paper explicitly notes that such a framework can serve as the trust and evidence layer beneath an “AgentRec” system that recommends the right agent for a task (Chishti et al., 30 Apr 2026). This is particularly relevant where agent selection depends not just on semantic fit but on verification strength, recency, and integrity signals.
6. Limitations, controversies, and open directions
Across its different meanings, AgentRec remains an unsettled research area. The SBERT-based routing paper is fast and accurate on its synthetic benchmark, but it relies on single-sentence synthetic prompts, does not provide full RLHF details, and does not test the claimed adaptability to new agents through an explicit class-expansion experiment (Park et al., 23 Jan 2025). The multi-agent conversational AgentRec paper offers a richer architecture and stronger benchmark section, but omits crucial implementation details such as backbone models, training objectives, prompt templates, and ablations, making it difficult to attribute gains to the coordinator, the specialized agents, or the tiered routing policy (Ma et al., 2 Oct 2025).
Adjacent work shows that these limitations are structural rather than accidental. “AgentRecBench” reports that strong agentic workflows can outperform simple baselines on some datasets while remaining weak on others, especially Yelp and certain evolving-interest settings, and it does not provide error bars or statistical significance tests (Shang et al., 26 May 2025). “AgenticTagger” shows that deeper levels of a hierarchical descriptor ontology become fuzzy and hard to disentangle, and that collaborative filtering signals are still absent from its vocabulary-building stage (Xie et al., 5 Feb 2026). “AgentReputation” leaves unresolved how to build verification ontologies, quantify verification strength, preserve privacy, bootstrap cold-start agents, and defend against collusion and other adversarial manipulations in decentralized markets (Chishti et al., 30 Apr 2026).
A broader implication is that AgentRec cannot be reduced to a single nearest-neighbor router, a single conversational architecture, or a single benchmark. The term now spans semantic retrieval, collaborative ranking, enterprise architecture selection, user-side mediation, reputation-aware allocation, and benchmark design. What unifies these strands is the need to convert heterogeneous context—task semantics, user intent, agent capability, interaction history, verification strength, and policy constraints—into a decision about which agent, item, or architecture should be selected next. The research trajectory suggests that future AgentRec systems will need tighter integration of representation learning, structured state, policy-aware trust, and reproducible evaluation, rather than incremental improvements to any one routing heuristic alone (Bogavelli et al., 13 Sep 2025).
