Toward Autonomous and Self-Evolving E-Commerce: A Review of Multi-Agent Personalization, Customer Digital Twins, Constrained Generative Interfaces, and Safe Policy Learning

Abstract

Personalization in electronic commerce is commonly implemented as a collection of separately optimized services for recommendation, search, advertising, and content selection. Such systems can identify relevant products while still presenting them through an unsuitable interface, an inappropriate explanation, or an action that conflicts with accessibility, privacy, inventory, or business constraints. This review examines the convergence of four research directions: behaviour-aware customer modelling, large-language-model-based multi-agent systems, constrained generative user interfaces, and reinforcement-learning-based policy improvement. On the basis of this synthesis, the paper proposes a reference architecture for an autonomous, self-evolving e-commerce platform. The architecture maintains a purpose-limited customer digital twin, infers session intent with calibrated uncertainty, collects typed proposals from specialist agents, and produces a declarative page specification over an approved component library. A deterministic policy layer validates every proposed adaptation before rendering. Behavioural and explicit preference signals are combined through recency-, relevance-, and confidence-aware scoring rather than a fixed event hierarchy. Learning is deliberately offline-first: candidate policies are evaluated from logged or simulated trajectories and promoted only after safety, accessibility, latency, and stability checks. The paper further defines research questions, baselines, metrics, and an ablation plan. The resulting framework positions generative AI as a bounded decision-support mechanism rather than an unrestricted source-code generator, thereby preserving auditability and operational control.