Ajay Upadhyay

Browser agents that ask a model what to do on every click are slow, expensive and unreliable at volume. On our test task, a computer-use agent needed 27 model calls, $0.30 and almost 7 minutes per row, and completed 7 rows out of 10. Teach-Once flips that. You do the task once in your browser while narrating it. Muse Voice Transcribe, Muse Spark and Muse Wonder compile the recording into a deterministic, parameterized workflow (plus a standalone Playwright script), which runs over a CSV file in about a second per row with no model calls at all. When the site changes and a step breaks, Muse Spark looks at the new page and works out what changed (moved, a different kind of control, split into parts, behind a Next button, removed, a popup in the way), and SAM 3.1 confirms exactly where the element is. The patch is verified, saved as a new workflow version, and reused by every later row, again with zero model calls. When a change can't be healed (blocked access, a new mandatory field the data doesn't have), the run stops and says why after one model call. It is not tied to one site: teach_live.py records on any website. On Demoland Hire, a modern React recruiting app included in the repo, three tasks taught by hand (add a candidate; search and remove one; schedule an interview through a three-step wizard with radio cards, a date, a custom dropdown and a switch) each compiled once ($0.16-0.43) and ran every CSV row with zero model calls, 1-2 s per row. With the app's own redesign switched on mid-run (new look, cookie banner, renamed menu and buttons, a dropdown turned into radio cards), the interview task healed on one row and finished 8 of 8 rows: 6 heals, $0.076 in total. Results on the demo portal: 50/50 rows through a mid-run redesign for about $0.04 in heals; 20 of 20 random redesigns recovered (two-page forms, custom dropdowns, removed fields, popups) at about $0.055 each; at 10,000 rows a day, roughly $0.30 once to compile versus about $3,000 a day for the agent.