A new arXiv paper targets a gap in browser agents: many systems perform well on short demonstrations but struggle on live websites where tasks require dozens of decisions.
Wuying-Browser-Agent is presented as a full pipeline rather than a single model trick. It includes a structured browser harness for stable execution, decision-oriented context management, supervised training on recovery trajectories and complex user interfaces, and an online reinforcement-learning method for long-horizon credit assignment.
The authors also introduce BrowserBench, a bilingual benchmark built around real web tasks. That matters because browser agents fail in ways that simple accuracy scores often hide: they get lost, click the wrong element, recover poorly, or succeed only on simplified pages.
The paper’s main claim is that deployment-quality browser agents need alignment across execution, supervision, optimization, and evaluation. It does not prove that general web automation is solved, but it gives researchers a more realistic target than clean, short scripted demos.