{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/f25b7c4bcb014a769b1d8cba8153b5e2\" frameborder=\"0\" width=\"1670\" height=\"1252\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1252,"width":1670,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1252,"thumbnail_width":1670,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/f25b7c4bcb014a769b1d8cba8153b5e2-45fbf59c14ce5dbe.gif","duration":188.365,"title":"Helping Agents Learn From Website Mistakes","description":"This Loom explains a “Memory for Computer Use” system that helps agents recover from mistakes during complex web and app tasks. It shows an agent struggling on Kayak when a flight search triggers a pop-up it cannot close, then describes how the system learns the correct response for that situation. The workflow uses Nimble agent search and Nimble Extract to choose sites and gather information, then runs a small on-device Vision language model (Liquid) to take actions in a loop (for example selecting calendar dates). After each task, traces are saved to two Route 3 database tables (traces and learnings), and a follow-up agent generates new learnings so success rate improves and task completion time decreases after a few runs."}