{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/72efd49f23ee478c8962e72c7ede7084\" frameborder=\"0\" width=\"2096\" height=\"1572\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1572,"width":2096,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1572,"thumbnail_width":2096,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/72efd49f23ee478c8962e72c7ede7084-6cdfa1c9de029f84.gif","duration":268.428,"title":"Autobrowse Scraping for Faster Job Extraction — AI agents that learn websites","description":"This Loom explains an autobrowse approach for AI scraping, where an agent discovers how to extract data from a site and then future runs skip the slow browser path. The author compares a naive Playwright-based extraction from Hacker News Who’s Hiring monthly threads (326 structured job postings in about 6 minutes for roughly 43 cents) against a learn phase that identifies a public Firebase API. By using a skill.md file with the discovered endpoints, schema, and parsing rules, the graduated path reduces API calls by 87% and cuts wall clock time by 44% from 6 minutes 38 seconds to about 3 minutes 43 seconds while keeping 99.7% extraction accuracy. They also note that every LLM call logs JSON telemetry for measurable cost and that the architecture is vendor agnostic via a single wrapper, with ongoing architecture decision records and retrospectives."}