Browser Use Open-Sources jev-ultrafast, Turning Controls into Numbered Lists

In September 2026, the Browser Use team open-sourced jev-ultrafast v0.1.0, a project aimed at a specific weakness in browser-controlling AI agents. As reported by DEV Community on October 3, 2026, the problem shows up in ordinary tasks such as asking an AI to search for flights: at every step the model takes a screenshot, reads the image, thinks, and then clicks. A single misread button means the whole step has to be redone, and by the time the agent finishes, the user could have found the flight on their own. According to the report, the method itself is the problem, and jev-ultrafast is built to attack exactly that.

Instead of feeding screenshots to the model by default, jev-ultrafast v0.1.0 lists the controls visible on the page as a numbered inventory. A separate decision model then picks which of those numbered items to act on next. The project says this design is meant to cut down on cases where the model misreads a button, and on the full-step restarts that follow from such mistakes.

That is the extent of what the coverage confirms so far: the open-source release in September 2026, the project name and version number, the shift away from screenshot input, and the failure mode it targets. The report offers no benchmark results or comparisons, no responses from other teams, and no word on whether further versions have followed v0.1.0.