Deploying a conversational shopping agent at production scale remains challenging despite rapid advances in large language models. Unlike research prototypes, production systems must satisfy strict requirements on accuracy, latency, reliability, and cost while serving millions of customers. We share our year-long journey of building a production shopping agent at scale and show that combining agentic reasoning with production-aware retrieval, tool orchestration, and system optimization enables a single shopping agent to support product discovery, shopping question answering, and agentic actions under real-world traffic. Our experience shows that LLM-based agents can be reliably operated at Amazon scale and provides practical design principles for industrial search and recommendation systems.

Modern e-commerce recommendation systems aim to improve customer experience by ranking content on search results page (SRP). However, displaying content is not always beneficial for customers across all contexts; even top-ranked content can be irrelevant, misleading, or redundant in certain scenarios. In this work, we propose a robust content suppression mechanism to selectively suppress content when necessary. Our approach leverages causal effect learning to measure the incremental value of showing versus suppressing content. Prior work uses Conditional Average Treatment Effect (CATE) to estimate treatment effect without considering the inherent uncertainty in uplift predictions, resulting in over-suppression and degraded customer experience. Our work introduces a novel Conformal Confidence (CoCo) suppressor that explicitly accounts for prediction uncertainty in uplift-based modeling. We first evaluate it on synthetic datasets with controlled noise, bias, and contexts. Results demonstrate superior performance compared to baseline approaches. Subsequently, our online traffic tests show statistically significant improvements in revenue and profit compared to existing methods.

E-commerce Search Results Pages (SRPs) are evolving from linear lists to complex, non-linear layouts, rendering traditional position-biased ranking models insufficient. Moreover, existing optimization frameworks typically maximize short-term signals (e.g., clicks, same-day revenue) because long-term satisfaction metrics (e.g., expected two-week revenue) involve delayed feedback and challenging long-horizon credit attribution. To bridge these gaps, we propose a novel Whole-Page Experience Optimization Framework. Unlike traditional list-wise rankers, our approach explicitly models the interplay between item relevance, 2D positional layout, and visual elements. We use a causal framework to develop metrics for measuring long-term user satisfaction based on quasi-experimental data. We validate our approach through industry-scale A/B testing, where the model demonstrated a 1.86% improvement in brand relevance (our primary customer experience metric) while simultaneously achieving a statistically significant revenue uplift of +0.05%.

Bandit algorithms arose as a standard approach to learning better models online. As they become more popular, they are increasingly deployed in complex machine learning pipelines, where their actions can be overwritten. For example, in ranking problems, a list of recommended items can be modified by a downstream algorithm to increase diversity. This may break the classic bandit algorithms and lead to linear regret. Specifically, if the proposed action is not taken, uncertainty in its estimated mean reward may not get reduced. In this work, we study this setting and call it non-compliant bandits; as the agent tries to learn rewarding actions that comply with a downstream task. We propose two algorithms, compliant contextual UCB (CompUCB) and Thompson sampling (CompTS), which learn separate reward and compliance models. The compliance model allows the agent to avoid non-compliant actions. We derive a sublinear regret bound for CompUCB. We also conduct experiments that compare our algorithms to classic bandit baselines. The experiments show failures of the baselines and that we mitigate them by learning compliance models.

Optimization is commonly employed to determine the content of web pages, such as to maximize conversions on landing pages or click-through rates on search engine result pages. Often the layout of these pages can be decoupled into several separate decisions. For example, the composition of a landing page may involve deciding which image to show, which wording to use, what color background to display, etc. Such optimization is a combinatorial problem over an exponentially large decision space. Randomized experiments do not scale well to this setting, and therefore, in practice, one is typically limited to optimizing a single aspect of a web page at a time. This represents a missed opportunity in both the speed of experimentation and the exploitation of possible interactions between layout decisions