SIGIR 2026 edition. This is the SIGIR 2026 edition of the tutorial. See also the RecSys 2025 edition.
When & where
| Detail | Information |
|---|---|
| Conference | SIGIR 2026, the 49th International ACM SIGIR Conference |
| Date | Monday, 20 July 2026 |
| Location | Melbourne Convention and Exhibition Centre, Melbourne, Australia |
| Room | Eureka 3 |
| Session time | 9:00 am to 12:00 pm (local time) |
| Official listing | SIGIR 2026 accepted tutorials |
About the tutorial
This tutorial covers multi-agentic recommender systems: recommender systems augmented with large language models (LLMs) and multi-agent orchestration to enable multi-step reasoning, tool use, and interactive decision-making. It emphasizes foundational concepts, reusable design patterns, and practical lessons learned from large-scale e-commerce deployments.
We begin with background and recent trends in generative recommender systems and their connection to agentic approaches. We then survey major deployment areas in industry and review the agent-orchestration frameworks developed to support them. Finally, we walk through a project that traces the full lifecycle of an agentic recommender system, from scoping and data definition through modeling, deployment, and monitoring, to provide actionable deployment insights.
The tutorial brings together perspectives from information retrieval (IR), recommender systems (RecSys), and large-scale industrial practice, and aims to give attendees practical insight into how these paradigms shape the design of next-generation user-facing IR and RecSys deployments.
What we’ll cover
The tutorial runs from foundations through to practice, drawing on what we’ve learned building and deploying these systems at scale. In particular:
- The move from generative to agentic recsys, with working definitions, levels of agent autonomy, and a survey of the key work in the area.
- Four major use cases for agents in recommendation, drawn from real search and personalization deployments, plus a look at the current agentic recsys startups.
- Memory for agents: ontology-driven memory, and retrieving from long, tool-heavy trajectories without being derailed by context noise.
- Keeping self-improving evolution agents from high variance and loops, and getting feature-engineering agents to describe items consistently.
- Harness and RL-based design that specifies the “what” rather than the “how,” and the practical side of putting agentic recsys into production.
- From open-source multi-agent frameworks to a single personal agent: one kernel adapting into many roles, coordinating a growing fleet across many owners, and a six-axis framework for characterizing agents.
- Evaluation metrics for agentic pipelines used in industry, with sample code where time allows.
Presenters
| Presenter | Affiliation |
|---|---|
| Reza Yousefi Maragheh | Walmart Global Tech |
| Yashar Deldjoo | Polytechnic University of Bari |
| Benjamin Coleman | Google DeepMind |
| Jason Cho | PreTask AI |
| Chi Wang | AG2AI |
Full presenter bios: TBD.
Who this is for
This tutorial is aimed at SIGIR attendees working on IR, recommendation, and LLM or agentic systems. Familiarity with basic RecSys and IR concepts is assumed. It is well suited for:
- Researchers and PhD students exploring agentic and multi-agent systems for recommendation and retrieval.
- Senior researchers and practitioners working with generative and LLM-based RecSys and IR.
- Industry teams looking for practical patterns and lessons to move from prototypes to reliable, large-scale deployments.
Materials
Slide decks are posted below as they become available. Remaining parts are TBD.
Part 2: Formalism and Lessons from Large Scale Deployments
Presented by Reza Yousefi Maragheh.
Part 3: Harness Engineering in Recommendation Systems Implementations and Practical Hurdles in Deploying Agents
Presented by Jason Cho.
Part 4: Self-Evolving Recommendation Systems
Presented by Benjamin Coleman.
Part 5: AgentOS Evolution
Presented by Chi Wang.
Other materials
- Supplementary material (planned): sample pseudocode for the different stages of productionizing a use case; example prompts, agent-role templates, and orchestration skeletons; and, where feasible, small toy datasets or public data excerpts suitable for demonstration.
- Curated reading list (planned): a survey plus representative papers to help attendees continue learning after the tutorial.
The accompanying material for the tutorial series can be found at agenticrecsys.github.io.
Citation
If you find this tutorial useful in your research or work, feel free to cite it:
@inproceedings{yousefi2026multi,
title={Multi-Agentic Recommender Systems: Foundations, Perspectives, and Lessons from Large Scale Deployments in eCommerce},
author={Yousefi Maragheh, Reza and Deldjoo, Yashar and Coleman, Benjamin and Cho, Jason and Wang, Chi},
booktitle={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages={5354--5356},
year={2026}
}
Yousefi Maragheh, R., Deldjoo, Y., Coleman, B., Cho, J., & Wang, C. (2026). Multi-Agentic Recommender Systems: Foundations, Perspectives, and Lessons from Large Scale Deployments in eCommerce. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 5354–5356). ACM.