Introduction to MACRS and the Design Challenge

Problem space: Goal-directed dialogue in conversational recommender systems (CRS)

Understanding the problem

Recommender systems play a central role in personalized digital experiences. Whether we’re choosing a movie, a product, or a restaurant, recommendations shape our decisions, often without us realizing it. But as user expectations evolve, traditional recommendation interfaces (sliders, forms, static suggestions) are being replaced with conversational recommenders. This is where users interact with an AI assistant via natural dialogue.

At first glance, large language models (LLMs) seem ideal for this job. They’re fluent, versatile, and capable of generating coherent conversations. Many recent approaches have proposed LLM-only conversational recommender systems (CRSs). These rely on a single model to handle everything, from asking for preferences to suggesting items and chit-chatting to keep the user engaged.

However, this LLM-only architecture has a hidden flaw.

Unlike open-ended chatbots, a CRS has a clear objective: gather enough preference data, find relevant recommendations, and present them persuasively, all while keeping the user engaged. That means balancing multiple skills simultaneously: information elicitation, recommendation, conversation management, and strategy refinement.

Traditional CRSs (pre-LLM) achieve this through hard-coded pipelines and explicit state tracking, often relying on slot-filling or finite-state machines. These systems could guide users toward a goal, but their interactions feel rigid and brittle. If a user says something unexpected or nuanced, the system will likely fail to adapt.

LLM-based systems, by contrast, can improvise, but they often lack structure and consistency in planning, memory, and strategic control. This leads to dialogue that might sound good, but that fails to make progress, loses track of user preferences, or repeats itself.

This is where MACRS comes in: a Multi-Agent Conversational Recommender System that merges the strengths of both worlds; the strategic planning of traditional systems and the generative fluency of LLMs.

Next, we’ll dive into what MACRS is, why it’s structured as a multi-agent system, and how it fundamentally improves the control and outcome of dialogue-based recommendations.

Introducing MACRS

To overcome the trade-off between fluency and control, MACRS (Multi-Agent Conversational Recommender System)Fang, Jiabao, Shen Gao, Pengjie Ren, Xiuying Chen, Suzan Verberne, and Zhaochun Ren. "A multi-agent conversational recommender system." arXiv preprint arXiv:2402.01135 (2024). proposes a different paradigm. Instead of relying on a single general-purpose agent to juggle all aspects of conversation, MACRS employs a team of specialized agents that collaborate to manage dialogue more strategically.

This design doesn’t just replicate what traditional CRSs do. It reimagines the architecture using agentic principles: assigning roles, aligning goals, coordinating behavior, and learning from experience. Specifically, MACRS exemplifies a modular agentic architecture that leverages an explicit control loop for dialogue flow, and addresses the inherent coordination costs of multi-agent systems through structured planning and reflection.

But before we dive into how it works, let’s reflect on the core design question that led to this system.

Design dilemma: One agent or many?

If you were to build your own agent-based conversational recommender system, how would you structure it?

You could take the single-agent route, feeding the entire dialogue history into an LLM and prompting it to decide what to do. It could ask more questions, recommend something, or keep chatting to build rapport. This is what many LLM-based CRSs do today.

Or you could adopt a multi-agent architecture, where different agents play different roles. One gathers preferences, another proposes items, and a third keeps the conversation natural and engaging. A central coordinator, yet another agent, decides which response to surface.

Here’s the catch: the first approach is simpler but brittle when the task gets complex. The second approach is more structured, modular, and robust, but it requires careful orchestration.

Which would you choose?

MACRS chose the latter. And for good reason.

Why go multi-agent?

Let’s unpack the rationale.

The challenge in CRS design is not just linguistic fluency; it’s dialogue control. The system needs to reason over:

• What the user already said.

• What information is still missing.

• Which strategy to follow (probe more, recommend now, or delay).

• How to keep the user engaged, even when no immediate recommendation is possible.

Trying to encode all of this into a single model’s prompt or context window quickly becomes unmanageable. Instead, MACRS distributes these competing conversational goals across agents with complementary expertise.

MACRS reframes the architecture of conversational recommendation from a monolithic agent to a collaborative system of specialists. It draws on a core principle of agentic design: distribute responsibility, but coordinate toward a shared goal.

Design goals for conversational recommendation

Imagine you’re chatting with a Smart Assistant. What would make that conversation helpful? That’s exactly what MACRS is designed to get right. To achieve that, it needs to:

1. Switch conversation styles: Sometimes the system needs to ask questions (“Do you like action movies?”), sometimes recommend something (“You might enjoy Inception”), and sometimes just keep the chat friendly (“That’s a great choice!”). A good assistant knows how to blend all three naturally.

2. Learn from behavior: Even if you don’t say “I don’t like horror,” skipping scary movie suggestions sends a signal. MACRS should pick up on that and adjust its strategy.

3. Decide what to say: One part of the system might want to ask more questions, another might be ready to recommend. MACRS has to choose the smartest next move.

4. Remember what matters: If you said you like comedy earlier or searched for romantic movies last week, the system should keep that in mind during the conversation.

5. Keep the conversation on track: The system isn’t just here to chat. It’s trying to help you find something you’ll actually choose. It needs to guide the conversation toward that goal, step-by-step.

These are the real-world challenges MACRS is built to handle. Next, we will learn how its design, built around multiple specialized agents working together, tackles each of these needs in a smart and coordinated way.

An overview of the MACRS system architecture

To effectively address the challenges of goal-directed dialogue, MACRS is built upon a sophisticated multi-agent architecture. As shown in the illustration below, the system comprises two essential, dynamically interacting modules:

1. Multi-agent act planning framework: This module focuses on controlling the dialogue flow by determining the most suitable conversational act at each turn. It involves a collaborative effort between specialized agents to generate and select responses.

2. User feedback-aware reflection mechanism: This dynamic optimization module allows MACRS to learn and adapt its behavior based on user interactions, continually refining its understanding and strategy.

The interaction between these modules forms a continuous improvement cycle. The multi-agent act planning workflow generates a response. The user reacts to it. The reflection mechanism observes this reaction (e.g., ‘user accepted the recommendation’ or ‘user skipped the suggestion’) and translates it into an update for the planning strategy and responder agents. This ensures the agent’s behavior adapts based on real-world outcomes and user feedback. This iterative process allows MACRS to continuously optimize its dialogue flow and recommendation accuracy.

We will delve into the detailed design and interaction of these components in the upcoming lessons of this case study. We will explore how the multi-agent act planning framework guides dialogue, and how the user feedback-aware reflection mechanism enables adaptation.