What is MiniWoB++?
MiniWoB++ is an open-source library of over 100 simulated, low-level web-interaction environments: clicking buttons, filling forms, using calendars and date pickers. Teams use it to train and benchmark reinforcement-learning and LLM-based web agents. It has a JavaScript interface where tasks render as real HTML/JS pages in a browser, and a Python interface built on Selenium WebDriver that follows the Gymnasium API. Observations cover DOM, utterance, fields, and screenshot data, with a reward signal per task. It extends the original OpenAI MiniWoB benchmark and is now maintained by the Farama Foundation, which also stewards Gymnasium and PettingZoo.
Our verdict
MiniWoB++ fits teams training or evaluating RL or LLM agents on structured, low-level web interactions: its Selenium-based Python interface and Gymnasium API drop into existing RL pipelines without much adaptation. Its tasks are simpler and more synthetic than open-web benchmarks like WebArena or Mind2Web, so a strong MiniWoB++ score doesn't tell you much about how an agent handles messy, real-world sites.
Categories:
These are research benchmarks, not products you buy. They define fixed web tasks and score how reliably an agent or model completes them, which is the closest thing the field has to an objective measure of web-agent capability. They are listed here for evaluation, not procurement.
