When web agents try to learn long tasks like booking a flight, they rarely succeed and get little feedback from their mistakes. This makes it hard to improve from experience alone.
SCOPE breaks the task into smaller skills, like clicking a date or entering a destination. It learns from both successes and failures, slowly combining skills into larger chains until the whole task is solved. Finally, it uses its own practice runs to internalize the full sequence.
Tested on three web benchmarks, SCOPE improved task success by 3.9% to 18.6% over the best prior methods. This shows that learning in small steps from failures is a practical way to train web agents.