Kernel Sphere is a DOM-native automation platform for developers which is built on Google Gemini and Playwright. You describe a task in plain English and give it a starting URL. The agent opens a real browser, reads the page, decides what to do at each step, and returns a structured result when it's done. It's open source. If you build on top of it, run into something it can't handle, or want to improve the extraction logic, contributions are welcome. github.com/kernelsphere-ai/kernelsphere

Building kernelsphere taught me a lot about where AI agents still struggle. Login flows, CAPTCHAs, multi-step forms, and sites that heavily rely on JavaScript all have their own failure modes. A lot of the work went into handling those edge cases rather than the happy path. We tested it on WebVoyager (643 tasks across 15 real websites) and hit 48% accuracy. There's a meaningful room to improve, and that's a big reason it's open source. Would love to hear from anyone who's worked on similar problems: What failure modes have you hit with browser agents that were hardest to fix? Are there specific site types or task categories you'd want to see it handle? GitHub is linked if you want to dig into the code or try it on your own tasks. Happy to answer anything about how it works.

Kernel Sphere

A web agent that reads, reasons, and acts in a real browser.

A web agent that reads, reasons, and acts in a real browser.