Browser UseFrom the builder
How we Made Scalable Long-Horizon RL Environments for Browser Use
Browser Use described how it builds reinforcement learning environments for browser agents from real user traffic against the live web, open sourcing 200 tasks with their rubrics as a sample. The team merged several hundred million production submissions into about ten million distinct tasks and screened them with Gemini 3 Flash, which matched human labels on 81 percent of a sealed test split and marked about 13 percent as suitable. Each task's rubric is written from the full set of attempts and checked against the live site, and after revision its items flip verdicts less than one percent of the time when frozen traces are regraded.
Browser agent benchmarks are saturating. Task sets that separated models a year ago now return pass rates above 90 percent.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.