Buy For Me — Agentic Shopping System
Buy For Me is an agentic commerce system within the Amazon Shopping app that enables customers to discover and purchase products from third-party brand websites when those items are not sold directly by Amazon. The system allows customers to delegate complex, multi-step shopping workflows, such as product discovery, configuration, and checkout, to an autonomous agent acting on their behalf, while remaining within the Amazon shopping experience. Unlike traditional integrations or chat-based assistants, Buy For Me operates directly on real third-party webpages, navigating dynamic user interfaces to extract information, make decisions, and execute actions.
Buy For Me autonomously executes end-to-end workflows that mirror how a human would shop on a third-party site, with additional verification and safety checks at each stage. During pre-purchase, the system surfaces eligible off-Amazon products and normalizes key product attributes so customers can evaluate options directly within the Amazon Shopping app. Once a customer authorizes a purchase, the agent executes the transaction by navigating the brand’s website on the customer’s behalf, managing multi-step interactions such as variant selection, cart updates, and checkout while handling dynamic UI behavior. After completion, post-purchase confirmations are ingested back into the app, enabling order tracking and support visibility while fulfillment, returns, and customer service remain managed by the brand.
Buy For Me is implemented as a modular agentic architecture that combines multimodal large language models for reasoning and planning with browser-based automation for executing actions on real third-party websites. Multimodal LLMs reason over page structure, rendered content, and user intent to plan multi-step workflows, while an action layer navigates dynamic UIs to interact with the websites and perform actions. A key challenge was achieving reliable, low-latency execution at scale across highly variable and partially observable web environments under strict product, safety, and trust constraints.
As a founding scientist behind Buy For Me, I contributed from early problem framing through productionization. My work spanned defining the end-to-end agentic workflow for off-Amazon purchasing, evaluating and selecting between alternative agent frameworks, and shaping the decision to build a first-party agentic system. I worked on the core system design, evaluation, data, and scaling strategy for purchase automation, built early proof-of-concept agents, authored key components of the multi-agent framework used for reasoning and action execution, and developed browser-level authentication mechanisms to enable automated, end-to-end purchasing.
Buy For Me has been covered by Forbes, Engadget, The Verge, 9to5mac and others.
Patents
- Systems for automated interaction with user interfaces (patent application)