WebShop

Why we included this project

Most language agents are trained on text, but WebShop drops them into a simulated storefront where they have to actually shop. The environment reproduces an e-commerce site with more than a million real product listings and roughly twelve thousand crowd-sourced instructions, so an agent must parse compositional requests, reformulate its own search queries, and wade through noisy page text to find and buy the right item. You can drive it through a browser-rendered HTML view or a lightweight OpenAI-gym text interface, which makes plugging in your own policy easy, and the repo ships rule-based, imitation-learning, and reinforcement-learning baselines for comparison. If you study web navigation, grounded language understanding, or decision-making in interactive settings, this is a solid, reproducible starting point, though it is a research tool rather than something you would deploy in production.