v.0.1.0
Super excited for the first version of Banana-lyzer, an open source AI Agent evaluation framework and dataset for web tasks with Playwright (And has a banana theme because why not) 🍌
We aim to solve the following issues with testing web agents:
- Websites change overtime, are affected by latency, and may have anti bot protections.
We need a system that can reliably save and deploy historic/static snapshots of websites. - Standard web practices are loose and there is an abundance of different underlying ways to represent a single individual website. For an agent to best generalize, we require building a diverse dataset of websites across industries and use-cases.
- We have specific evaluation criteria and agent use cases focusing on structured and direct information retrieval across websites.
- There exists valuable web task datasets and evaluations that we'd like to unify in a single repo (Mind2Web, WebArena, etc).